{
 "source": "https://vramcalculator.net",
 "updated": "2026-10-14",
 "license": "Free to use with attribution to VRAMCalculator.net",
 "items": [
  {
   "slug": "llama-3-2-1b",
   "name": "Llama 3.2 1B",
   "family": "Llama",
   "org": "Meta",
   "params": 1.24,
   "layers": 16,
   "hidden": 2048,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 64,
   "vocab": 128256,
   "inter": 8192,
   "maxCtx": 131072,
   "released": "2024-09",
   "license": "Llama 3.2 Community License",
   "hf": "meta-llama/Llama-3.2-1B-Instruct",
   "notes": "Meta's smallest Llama 3 model, built for phones and edge devices. The 128K-token vocabulary is a large share of its 1.24B parameters, so it quantizes less efficiently than its size suggests. Useful as a draft model for speculative decoding with larger Llama models.",
   "faq": [
    {
     "q": "Can Llama 3.2 1B run without a GPU?",
     "a": "Yes. At Q4_K_M the weights are under 1 GB, so any modern CPU runs it at usable speed. A GPU only matters if you need very high throughput."
    }
   ]
  },
  {
   "slug": "llama-3-2-3b",
   "name": "Llama 3.2 3B",
   "family": "Llama",
   "org": "Meta",
   "params": 3.21,
   "layers": 28,
   "hidden": 3072,
   "heads": 24,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 128256,
   "inter": 8192,
   "maxCtx": 131072,
   "released": "2024-09",
   "license": "Llama 3.2 Community License",
   "hf": "meta-llama/Llama-3.2-3B-Instruct",
   "notes": "A compact Llama with the full 128K context window. Its KV cache grows faster than the 1B model's because it has 28 layers with 128-dimension heads, so long-context use can need more memory than the weights themselves.",
   "faq": [
    {
     "q": "Is Llama 3.2 3B good for a 4GB or 6GB GPU?",
     "a": "Yes. At Q4_K_M it needs about 3.5 GB with an 8K context, which leaves room for a longer context on 6 GB cards."
    }
   ]
  },
  {
   "slug": "llama-3-1-8b",
   "name": "Llama 3.1 8B",
   "family": "Llama",
   "org": "Meta",
   "params": 8.03,
   "layers": 32,
   "hidden": 4096,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 128256,
   "inter": 14336,
   "maxCtx": 131072,
   "released": "2024-07",
   "license": "Llama 3.1 Community License",
   "hf": "meta-llama/Llama-3.1-8B-Instruct",
   "notes": "The default local model for most people. Grouped-query attention with 8 KV heads keeps the cache at 128 KB per token in FP16, so a full 128K context adds 16 GB on top of the weights. Fine-tunes of this model are among the most common on Hugging Face.",
   "faq": [
    {
     "q": "Can Llama 3.1 8B run on an 8GB GPU?",
     "a": "Yes, at Q4_K_M with up to about 16K context. At Q8_0 the weights alone are about 8 GB, so an 8 GB card needs CPU offload."
    },
    {
     "q": "How much VRAM does Llama 3.1 8B need for 128K context?",
     "a": "About 16 GB for the KV cache alone in FP16, so roughly 21 GB total at Q4_K_M. Quantizing the cache to Q8_0 cuts that cache to about 8.5 GB."
    }
   ]
  },
  {
   "slug": "llama-3-3-70b",
   "name": "Llama 3.3 70B",
   "family": "Llama",
   "org": "Meta",
   "params": 70.6,
   "layers": 80,
   "hidden": 8192,
   "heads": 64,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 128256,
   "inter": 28672,
   "maxCtx": 131072,
   "released": "2024-12",
   "license": "Llama 3.3 Community License",
   "hf": "meta-llama/Llama-3.3-70B-Instruct",
   "notes": "Same architecture as Llama 3.1 70B with better post-training; Meta positioned it as close to Llama 3.1 405B on many benchmarks. The 80 layers make the KV cache 320 KB per token in FP16, so context length matters more here than on 8B models. The classic setup is two 24 GB cards at Q4_K_M.",
   "faq": [
    {
     "q": "Can Llama 3.3 70B run on a single RTX 4090?",
     "a": "Not fully. Q4_K_M needs about 43 GB with an 8K context. On 24 GB you need Q2_K (still over 24 GB) or heavy CPU offload, which drops speed to a few tokens per second."
    },
    {
     "q": "What is the cheapest way to run Llama 3.3 70B?",
     "a": "Two used RTX 3090s (48 GB total) at Q4_K_M, or a Mac or Ryzen AI Max system with 64 GB or more of unified memory."
    }
   ]
  },
  {
   "slug": "llama-3-1-405b",
   "name": "Llama 3.1 405B",
   "family": "Llama",
   "org": "Meta",
   "params": 405.9,
   "layers": 126,
   "hidden": 16384,
   "heads": 128,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 128256,
   "inter": 53248,
   "maxCtx": 131072,
   "released": "2024-07",
   "license": "Llama 3.1 Community License",
   "hf": "meta-llama/Llama-3.1-405B-Instruct",
   "notes": "Meta's largest dense open model. Even at Q4_K_M it needs well over 200 GB, which puts it in multi-GPU server or 512 GB Mac Studio territory. Meta also published an official FP8 version for 8x H100 serving.",
   "faq": [
    {
     "q": "Can Llama 3.1 405B run on a Mac Studio?",
     "a": "A 512 GB M3 Ultra can hold Q4_K_M with a moderate context once you raise the GPU memory limit, but expect only a few tokens per second because every token reads about 230 GB of weights."
    }
   ]
  },
  {
   "slug": "llama-4-scout",
   "name": "Llama 4 Scout",
   "family": "Llama",
   "org": "Meta",
   "params": 109,
   "active": 17,
   "layers": 48,
   "hidden": 5120,
   "heads": 40,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 202048,
   "maxCtx": 131072,
   "swa": {
    "window": 8192,
    "global": 12
   },
   "released": "2025-04",
   "license": "Llama 4 Community License",
   "hf": "meta-llama/Llama-4-Scout-17B-16E-Instruct",
   "notes": "A mixture-of-experts model with 16 experts: 109B total parameters, 17B active per token. All 109B must sit in memory, but speed behaves like a 17B model. Three of every four layers use chunked attention over 8K tokens, so the KV cache grows slowly with context. Meta advertises a 10M-token window; this calculator caps it at 128K.",
   "faq": [
    {
     "q": "Does Llama 4 Scout need VRAM for 17B or 109B parameters?",
     "a": "For 109B. Every expert must be loaded because any token can route to any expert. The 17B active figure only sets the speed."
    }
   ]
  },
  {
   "slug": "qwen-2-5-7b",
   "name": "Qwen 2.5 7B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 7.62,
   "layers": 28,
   "hidden": 3584,
   "heads": 28,
   "kvHeads": 4,
   "headDim": 128,
   "vocab": 152064,
   "inter": 18944,
   "maxCtx": 32768,
   "released": "2024-09",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen2.5-7B-Instruct",
   "notes": "Only 4 KV heads across 28 layers make this one of the most cache-efficient 7B models: 56 KB per token in FP16, less than half of Llama 3.1 8B. The base config allows 32K tokens; YaRN scaling extends it to 128K.",
   "faq": [
    {
     "q": "Is Qwen 2.5 7B lighter than Llama 3.1 8B?",
     "a": "Slightly on weights, and much lighter on context: its KV cache per token is under half the size, so it fits longer contexts on the same card."
    }
   ]
  },
  {
   "slug": "qwen-2-5-14b",
   "name": "Qwen 2.5 14B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 14.7,
   "layers": 48,
   "hidden": 5120,
   "heads": 40,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 152064,
   "inter": 13824,
   "maxCtx": 32768,
   "released": "2024-09",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen2.5-14B-Instruct",
   "notes": "A sweet spot for 12 GB and 16 GB cards. At Q4_K_M the weights are about 8.3 GB, leaving room for context on a 12 GB GPU. Its 48 layers give a KV cache of 192 KB per token in FP16.",
   "faq": [
    {
     "q": "Does Qwen 2.5 14B fit on an RTX 3060 12GB?",
     "a": "Yes at Q4_K_M with an 8K context (about 10.7 GB). For 32K context, quantize the KV cache to Q8_0 or use a 16 GB card."
    }
   ]
  },
  {
   "slug": "qwen-2-5-32b",
   "name": "Qwen 2.5 32B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 32.8,
   "layers": 64,
   "hidden": 5120,
   "heads": 40,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 152064,
   "inter": 27648,
   "maxCtx": 32768,
   "released": "2024-09",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen2.5-32B-Instruct",
   "notes": "The model that made 24 GB cards the standard for local AI. Q4_K_M weights are about 18.5 GB, which fits on an RTX 3090 or 4090 with roughly 12K of context. The same architecture underlies QwQ-32B and DeepSeek R1 Distill Qwen 32B.",
   "faq": [
    {
     "q": "Can Qwen 2.5 32B run on 16GB VRAM?",
     "a": "Only at Q3_K_M or lower with a short context, or with partial CPU offload. 24 GB is the comfortable minimum for Q4_K_M."
    }
   ]
  },
  {
   "slug": "qwen-2-5-coder-32b",
   "name": "Qwen 2.5 Coder 32B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 32.8,
   "layers": 64,
   "hidden": 5120,
   "heads": 40,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 152064,
   "inter": 27648,
   "maxCtx": 32768,
   "released": "2024-11",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen2.5-Coder-32B-Instruct",
   "notes": "The coding variant of Qwen 2.5 32B, with identical memory needs. Coding agents send long files and repository context, so plan VRAM for 32K tokens rather than 8K: that adds about 8 GB of FP16 cache.",
   "faq": [
    {
     "q": "What context length should I plan for with a coding model?",
     "a": "At least 32K for editor and agent use. For Qwen 2.5 Coder 32B that means about 27 GB at Q4_K_M with an FP16 cache, about 23.6 GB with a Q8_0 cache (a tight fit on 24 GB), or about 21.6 GB with a Q4_0 cache."
    }
   ]
  },
  {
   "slug": "qwen-2-5-72b",
   "name": "Qwen 2.5 72B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 72.7,
   "layers": 80,
   "hidden": 8192,
   "heads": 64,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 152064,
   "inter": 29568,
   "maxCtx": 32768,
   "released": "2024-09",
   "license": "Qwen License",
   "hf": "Qwen/Qwen2.5-72B-Instruct",
   "notes": "Qwen's largest dense 2.5 model, with the same layer and head layout as Llama 70B. Memory needs are slightly higher than Llama 3.3 70B because of the larger vocabulary and wider feed-forward layers.",
   "faq": [
    {
     "q": "Is Qwen 2.5 72B harder to run than Llama 3.3 70B?",
     "a": "Barely. It is about 1 GB larger at Q4_K_M and has the same KV cache size per token."
    }
   ]
  },
  {
   "slug": "qwen-3-4b",
   "name": "Qwen 3 4B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 4.02,
   "layers": 36,
   "hidden": 2560,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 151936,
   "inter": 9728,
   "maxCtx": 40960,
   "released": "2025-04",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3-4B",
   "notes": "A small hybrid thinking model. It has 36 layers, more than most 4B models, so its KV cache (144 KB per token in FP16) is larger than Llama 3.1 8B's. Reasoning traces run to thousands of tokens, so give it more context than a chat model needs.",
   "faq": [
    {
     "q": "Why does Qwen 3 4B use so much memory for context?",
     "a": "It is deep (36 layers) with 8 KV heads of 128 dimensions. That gives 144 KB per token in FP16, so 32K of context adds about 4.5 GB."
    }
   ]
  },
  {
   "slug": "qwen-3-8b",
   "name": "Qwen 3 8B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 8.19,
   "layers": 36,
   "hidden": 4096,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 151936,
   "inter": 12288,
   "maxCtx": 40960,
   "released": "2025-04",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3-8B",
   "notes": "Supports switching between thinking and non-thinking modes. With 36 layers its cache is 144 KB per token in FP16, about 12% more than Llama 3.1 8B. The native window is 32K (40,960 in config.json), extendable to 128K with YaRN.",
   "faq": [
    {
     "q": "Can Qwen 3 8B run on an 8GB GPU?",
     "a": "Yes at Q4_K_M with an 8K context (about 6.6 GB). Thinking mode can use long traces, so a Q8_0 KV cache helps on 8 GB cards."
    }
   ]
  },
  {
   "slug": "qwen-3-14b",
   "name": "Qwen 3 14B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 14.8,
   "layers": 40,
   "hidden": 5120,
   "heads": 40,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 151936,
   "inter": 17408,
   "maxCtx": 40960,
   "released": "2025-04",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3-14B",
   "notes": "A strong reasoning model for 12 GB and 16 GB cards. With 40 layers (fewer than Qwen 2.5 14B's 48) its KV cache is lighter at 160 KB per token in FP16.",
   "faq": [
    {
     "q": "Qwen 3 14B or Qwen 2.5 14B on a 12GB card?",
     "a": "Qwen 3 14B has slightly larger weights but a smaller KV cache, so at 16K context or more it actually needs less VRAM."
    }
   ]
  },
  {
   "slug": "qwen-3-32b",
   "name": "Qwen 3 32B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 32.8,
   "layers": 64,
   "hidden": 5120,
   "heads": 64,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 151936,
   "inter": 25600,
   "maxCtx": 40960,
   "released": "2025-04",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3-32B",
   "notes": "The largest dense Qwen 3 model. Weights match Qwen 2.5 32B in size, and the KV cache is the same 256 KB per token in FP16. A 24 GB card runs it at Q4_K_M with about 12K of context.",
   "faq": [
    {
     "q": "Is Qwen 3 32B or Qwen 3 30B-A3B better for a 24GB GPU?",
     "a": "30B-A3B is several times faster because only 3.3B parameters are active per token. 32B is the stronger model per token. Both fit 24 GB at Q4_K_M."
    }
   ]
  },
  {
   "slug": "qwen-3-30b-a3b",
   "name": "Qwen 3 30B-A3B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 30.5,
   "active": 3.3,
   "layers": 48,
   "hidden": 2048,
   "heads": 32,
   "kvHeads": 4,
   "headDim": 128,
   "vocab": 151936,
   "maxCtx": 40960,
   "released": "2025-04",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3-30B-A3B",
   "notes": "A mixture-of-experts model with 128 experts, 8 active per token. It needs memory for all 30.5B parameters but generates text at the speed of a 3B model. With only 4 KV heads the cache is light at 96 KB per token in FP16. Because so little is read per token, it stays usable even with partial CPU offload.",
   "faq": [
    {
     "q": "Why is Qwen 3 30B-A3B so fast?",
     "a": "Each token only reads the 3.3B active parameters, about 2 GB at Q4_K_M. Speed is limited by memory bandwidth, so it runs roughly as fast as a 3B dense model."
    },
    {
     "q": "Can Qwen 3 30B-A3B run on 16GB VRAM?",
     "a": "With partial offload, yes. The weights are about 17 GB at Q4_K_M; keeping the expert layers on the CPU still gives usable speed because each token touches few experts."
    }
   ]
  },
  {
   "slug": "qwen-3-235b-a22b",
   "name": "Qwen 3 235B-A22B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 235.1,
   "active": 22,
   "layers": 94,
   "hidden": 4096,
   "heads": 64,
   "kvHeads": 4,
   "headDim": 128,
   "vocab": 151936,
   "maxCtx": 40960,
   "released": "2025-04",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3-235B-A22B",
   "notes": "Qwen's flagship mixture-of-experts model. All 235B parameters must be in memory (about 133 GB at Q4_K_M), but only 22B are read per token. It suits 192 GB+ Macs, multi-GPU servers, or hybrid setups that keep the experts in system RAM.",
   "faq": [
    {
     "q": "What hardware runs Qwen 3 235B-A22B locally?",
     "a": "A 256 GB or 512 GB Mac Studio, four or more 48 GB GPUs, or one large GPU plus 192 GB+ of fast system RAM using expert offload in llama.cpp."
    }
   ]
  },
  {
   "slug": "qwen-3-5-4b",
   "name": "Qwen 3.5 4B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 4.2,
   "layers": 32,
   "hidden": 2560,
   "heads": 16,
   "kvHeads": 4,
   "headDim": 256,
   "vocab": 248320,
   "inter": 9216,
   "maxCtx": 262144,
   "linear": {
    "layers": 24,
    "kHeads": 16,
    "vHeads": 32,
    "dim": 128
   },
   "released": "2026-03",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3.5-4B",
   "notes": "The small member of the Qwen 3.5 family, with the same hybrid design as its big siblings: three Gated DeltaNet (linear attention) layers for every full-attention layer. Only 8 of its 32 layers keep a KV cache, so a 32K or even 128K context costs a fraction of what it does on Qwen 3 4B. It is natively multimodal; the vision encoder is a separate file (mmproj) in GGUF builds.",
   "faq": [
    {
     "q": "Is Qwen 3.5 4B better than Qwen 3 4B for low-VRAM GPUs?",
     "a": "For long contexts, yes. The weights are about the same size, but only a quarter of the layers store a KV cache, so 32K of context costs about a quarter of what Qwen 3 4B needs."
    }
   ]
  },
  {
   "slug": "qwen-3-5-9b",
   "name": "Qwen 3.5 9B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 8.95,
   "layers": 32,
   "hidden": 4096,
   "heads": 16,
   "kvHeads": 4,
   "headDim": 256,
   "vocab": 248320,
   "inter": 12288,
   "maxCtx": 262144,
   "linear": {
    "layers": 24,
    "kHeads": 16,
    "vHeads": 32,
    "dim": 128
   },
   "released": "2026-03",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3.5-9B",
   "notes": "A dense 9B model that replaces Qwen 3 8B for most local users. Hybrid attention means only 8 layers keep a KV cache, so it can use a 64K context on a 12 GB card. The 248K-token vocabulary adds about 2 GB of embedding weights in FP16, which is why it is slightly larger than its name suggests.",
   "faq": [
    {
     "q": "Can Qwen 3.5 9B run on an 8GB GPU?",
     "a": "Yes at Q4_K_M with a moderate context. The weights are about 5 GB, and because only 8 of 32 layers store a cache, even a 32K context adds only about 1 GB."
    }
   ]
  },
  {
   "slug": "qwen-3-5-27b",
   "name": "Qwen 3.5 27B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 26.9,
   "layers": 64,
   "hidden": 5120,
   "heads": 24,
   "kvHeads": 4,
   "headDim": 256,
   "vocab": 248320,
   "inter": 17408,
   "maxCtx": 262144,
   "linear": {
    "layers": 48,
    "kHeads": 16,
    "vHeads": 48,
    "dim": 128
   },
   "released": "2026-02",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3.5-27B",
   "notes": "The first dense Qwen with hybrid attention: 16 gated full-attention layers and 48 Gated DeltaNet layers. That makes long context unusually cheap for a model this size, and it is the reason a 24 GB card can hold it at Q4_K_M with a 64K+ context. Later releases (Qwen 3.6 27B and 3.8 27B) keep the exact same architecture, so memory figures are identical.",
   "faq": [
    {
     "q": "Does Qwen 3.5 27B fit on a 24GB GPU?",
     "a": "Yes. At Q4_K_M the weights are about 15 GB, and the KV cache only grows on 16 of 64 layers, so a 24 GB card has room for a long context."
    },
    {
     "q": "Why does my runtime not support Qwen 3.5?",
     "a": "Gated DeltaNet layers need runtime support. Use a recent llama.cpp, Ollama, LM Studio or vLLM build; older versions fail to load the model."
    }
   ]
  },
  {
   "slug": "qwen-3-5-35b-a3b",
   "name": "Qwen 3.5 35B-A3B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 35,
   "active": 3,
   "layers": 40,
   "hidden": 2048,
   "heads": 16,
   "kvHeads": 2,
   "headDim": 256,
   "vocab": 248320,
   "maxCtx": 262144,
   "linear": {
    "layers": 30,
    "kHeads": 16,
    "vHeads": 32,
    "dim": 128
   },
   "released": "2026-02",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3.5-35B-A3B",
   "notes": "The successor to Qwen 3 30B-A3B: 256 experts with 8 routed plus one shared expert per token, about 3B active. Only 10 of 40 layers keep a KV cache, with just 2 KV heads, so the cache is tiny. It is fast on almost anything that can hold the weights, and it stays usable with partial CPU offload.",
   "faq": [
    {
     "q": "Can Qwen 3.5 35B-A3B run on a 16GB GPU?",
     "a": "Not fully at Q4_K_M, which needs about 21 GB. Because only about 3B parameters are read per token, offloading the expert layers to system RAM still gives usable speed, and IQ3-level quants fit fully."
    }
   ]
  },
  {
   "slug": "qwen-3-5-122b-a10b",
   "name": "Qwen 3.5 122B-A10B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 122,
   "active": 10,
   "layers": 48,
   "hidden": 3072,
   "heads": 32,
   "kvHeads": 2,
   "headDim": 256,
   "vocab": 248320,
   "maxCtx": 262144,
   "linear": {
    "layers": 36,
    "kHeads": 16,
    "vHeads": 64,
    "dim": 128
   },
   "released": "2026-02",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3.5-122B-A10B",
   "notes": "A mid-size MoE aimed at 96–128 GB machines. With about 10B active parameters it generates text roughly as fast as a 10B dense model, while needing memory for all 122B. The hybrid attention leaves only 12 layers with a KV cache, so even a 128K context adds just a few gigabytes.",
   "faq": [
    {
     "q": "What hardware runs Qwen 3.5 122B-A10B locally?",
     "a": "At Q4_K_M it needs about 70 GB, so a 96 GB or 128 GB Mac, a Ryzen AI Max+ 395 system, DGX Spark, an RTX Pro 6000 or three to four 24 GB GPUs."
    }
   ]
  },
  {
   "slug": "qwen-3-5-397b-a17b",
   "name": "Qwen 3.5 397B-A17B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 397,
   "active": 17,
   "layers": 60,
   "hidden": 4096,
   "heads": 32,
   "kvHeads": 2,
   "headDim": 256,
   "vocab": 248320,
   "maxCtx": 262144,
   "linear": {
    "layers": 45,
    "kHeads": 16,
    "vHeads": 64,
    "dim": 128
   },
   "released": "2026-02",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3.5-397B-A17B",
   "notes": "Qwen's flagship open model of early 2026: 512 experts with 10 routed plus one shared per token. It is a server or Mac Studio model; at Q4_K_M it needs well over 200 GB. Its KV cache is remarkably small for its size because only 15 of 60 layers use full attention, each with 2 KV heads.",
   "faq": [
    {
     "q": "Can Qwen 3.5 397B run on a Mac Studio?",
     "a": "Yes on a 512 GB M3 Ultra, which holds Q4_K_M or even Q6_K with a long context. A 256 GB machine needs an IQ3 or smaller quant."
    }
   ]
  },
  {
   "slug": "qwen-3-6-27b",
   "name": "Qwen 3.6 27B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 26.9,
   "layers": 64,
   "hidden": 5120,
   "heads": 24,
   "kvHeads": 4,
   "headDim": 256,
   "vocab": 248320,
   "inter": 17408,
   "maxCtx": 262144,
   "linear": {
    "layers": 48,
    "kHeads": 16,
    "vHeads": 48,
    "dim": 128
   },
   "released": "2026-04",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3.6-27B",
   "notes": "A post-training refresh of Qwen 3.5 27B with stronger coding and agent skills. The config.json is identical to 3.5 27B, so any GGUF of the same quantization has the same size and the same memory needs. If you already run 3.5 27B, 3.6 drops into the same hardware.",
   "faq": [
    {
     "q": "Does Qwen 3.6 27B need more VRAM than Qwen 3.5 27B?",
     "a": "No. The architecture is identical, so the same quantization needs exactly the same memory."
    }
   ]
  },
  {
   "slug": "qwen-3-6-35b-a3b",
   "name": "Qwen 3.6 35B-A3B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 35,
   "active": 3,
   "layers": 40,
   "hidden": 2048,
   "heads": 16,
   "kvHeads": 2,
   "headDim": 256,
   "vocab": 248320,
   "maxCtx": 262144,
   "linear": {
    "layers": 30,
    "kHeads": 16,
    "vHeads": 32,
    "dim": 128
   },
   "released": "2026-04",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3.6-35B-A3B",
   "notes": "An updated Qwen 3.5 35B-A3B with the same architecture and better coding results. It is one of the most popular models for 24 GB cards and 32 GB Macs because it runs at small-model speed with a very small KV cache. Official FP8 weights are also published for vLLM users.",
   "faq": [
    {
     "q": "How fast is Qwen 3.6 35B-A3B?",
     "a": "It reads only about 3B parameters per token, so on a 24 GB GPU it generates text several times faster than a dense 27B model of similar quality."
    }
   ]
  },
  {
   "slug": "qwen-3-8-27b",
   "name": "Qwen 3.8 27B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 26.9,
   "layers": 64,
   "hidden": 5120,
   "heads": 24,
   "kvHeads": 4,
   "headDim": 256,
   "vocab": 248320,
   "inter": 17408,
   "maxCtx": 262144,
   "linear": {
    "layers": 48,
    "kHeads": 16,
    "vHeads": 48,
    "dim": 128
   },
   "released": "2026-08",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3.8-27B",
   "notes": "The latest dense Qwen at the 27B size, released in August 2026. It keeps the Qwen 3.5 hybrid layout (16 full-attention and 48 Gated DeltaNet layers) and a native 262K context, extendable toward 1M tokens with YaRN. Memory needs match Qwen 3.5 and 3.6 27B exactly; only quality changed.",
   "faq": [
    {
     "q": "How much context can Qwen 3.8 27B use on a 24GB GPU?",
     "a": "At Q4_K_M, about 96K tokens with an FP16 cache and about 180K with a Q8_0 cache. That is far more than a dense model without hybrid attention allows, because only 16 of 64 layers store keys and values."
    }
   ]
  },
  {
   "slug": "qwen-3-next-80b-a3b",
   "name": "Qwen3-Next 80B-A3B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 80,
   "active": 3,
   "layers": 48,
   "hidden": 2048,
   "heads": 16,
   "kvHeads": 2,
   "headDim": 256,
   "vocab": 151936,
   "maxCtx": 262144,
   "linear": {
    "layers": 36,
    "kHeads": 16,
    "vHeads": 32,
    "dim": 128
   },
   "released": "2025-09",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3-Next-80B-A3B-Instruct",
   "notes": "The experimental model that introduced Qwen's hybrid Gated DeltaNet plus gated attention design, later used across Qwen 3.5. It has 512 experts but only about 3B active parameters, so it is very fast once loaded. The full 80B must sit in memory, which makes it a good fit for 64 GB Macs and 48 GB GPU setups.",
   "faq": [
    {
     "q": "Is Qwen3-Next 80B-A3B good for a 64GB Mac?",
     "a": "Yes. Just. Q4_K_M needs about 46 GB, right at the roughly 48 GB a 64 GB Mac gives the GPU by default, so keep the context short or raise the wired-memory limit. Once loaded it generates at small-model speed."
    }
   ]
  },
  {
   "slug": "qwen-3-coder-30b-a3b",
   "name": "Qwen3-Coder 30B-A3B",
   "family": "Qwen",
   "org": "Alibaba",
   "params": 30.5,
   "active": 3.3,
   "layers": 48,
   "hidden": 2048,
   "heads": 32,
   "kvHeads": 4,
   "headDim": 128,
   "vocab": 151936,
   "maxCtx": 262144,
   "released": "2025-07",
   "license": "Apache 2.0",
   "hf": "Qwen/Qwen3-Coder-30B-A3B-Instruct",
   "notes": "The local-friendly coding model in the Qwen3-Coder line, built on the Qwen 3 30B-A3B architecture with a 256K native context. It is popular with coding agents such as Cline, Roo Code and Continue because it is fast. Agents use long contexts, so the KV cache, not the weights, is usually what runs out first.",
   "faq": [
    {
     "q": "How much VRAM does Qwen3-Coder 30B need for a 128K context?",
     "a": "About 12 GB of KV cache in FP16 on top of roughly 17 GB of Q4_K_M weights, so about 30 GB. A Q8_0 cache brings that to about 25 GB."
    }
   ]
  },
  {
   "slug": "mistral-7b",
   "name": "Mistral 7B v0.3",
   "family": "Mistral",
   "org": "Mistral AI",
   "params": 7.25,
   "layers": 32,
   "hidden": 4096,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 32768,
   "inter": 14336,
   "maxCtx": 32768,
   "released": "2024-05",
   "license": "Apache 2.0",
   "hf": "mistralai/Mistral-7B-Instruct-v0.3",
   "notes": "The model that popularized grouped-query attention in open 7B models. Its small 32K vocabulary means less memory goes to embeddings, so its weights are lighter than Llama 3.1 8B's at every quantization.",
   "faq": [
    {
     "q": "Is Mistral 7B still worth running?",
     "a": "It remains a common base for fine-tunes and is very light, but newer models of the same size usually score higher. Memory needs are nearly identical to Llama 3.1 8B."
    }
   ]
  },
  {
   "slug": "mistral-nemo-12b",
   "name": "Mistral Nemo 12B",
   "family": "Mistral",
   "org": "Mistral AI & NVIDIA",
   "params": 12.2,
   "layers": 40,
   "hidden": 5120,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 131072,
   "inter": 14336,
   "maxCtx": 131072,
   "released": "2024-07",
   "license": "Apache 2.0",
   "hf": "mistralai/Mistral-Nemo-Instruct-2407",
   "notes": "Built with NVIDIA and trained for FP8 inference without quality loss. It is popular for role-play and creative fine-tunes. With 40 layers the cache is 160 KB per token in FP16, so 128K context needs 20 GB on top of the weights.",
   "faq": [
    {
     "q": "What GPU does Mistral Nemo 12B need?",
     "a": "A 12 GB card runs Q4_K_M with up to about 20K context. 16 GB cards fit Q6_K or longer contexts."
    }
   ]
  },
  {
   "slug": "mistral-small-3-1-24b",
   "name": "Mistral Small 3.1 24B",
   "family": "Mistral",
   "org": "Mistral AI",
   "params": 24,
   "layers": 40,
   "hidden": 5120,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 131072,
   "inter": 32768,
   "maxCtx": 131072,
   "released": "2025-03",
   "license": "Apache 2.0",
   "hf": "mistralai/Mistral-Small-3.1-24B-Instruct-2503",
   "notes": "A dense 24B model with vision input, designed to run on a single RTX 4090 or a 32 GB Mac. It has only 40 layers for its size, so its KV cache is modest: 160 KB per token in FP16.",
   "faq": [
    {
     "q": "Does Mistral Small 3.1 fit on a 16GB GPU?",
     "a": "At Q4_K_M the weights are about 13.6 GB, so a 16 GB card fits it only with a short context (about 3K tokens). 24 GB is comfortable."
    }
   ]
  },
  {
   "slug": "mixtral-8x7b",
   "name": "Mixtral 8x7B",
   "family": "Mistral",
   "org": "Mistral AI",
   "params": 46.7,
   "active": 12.9,
   "layers": 32,
   "hidden": 4096,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 32000,
   "maxCtx": 32768,
   "released": "2023-12",
   "license": "Apache 2.0",
   "hf": "mistralai/Mixtral-8x7B-Instruct-v0.1",
   "notes": "The first widely used open mixture-of-experts model: 8 experts, 2 active per token. Memory follows the 46.7B total while speed follows the 12.9B active parameters.",
   "faq": [
    {
     "q": "Why is Mixtral 8x7B 47B and not 56B?",
     "a": "Only the feed-forward layers are duplicated across the 8 experts. Attention and embeddings are shared, so the total is 46.7B."
    }
   ]
  },
  {
   "slug": "mistral-small-4",
   "name": "Mistral Small 4 119B",
   "family": "Mistral",
   "org": "Mistral AI",
   "params": 119,
   "active": 6.5,
   "layers": 36,
   "hidden": 4096,
   "heads": 32,
   "kvHeads": 32,
   "headDim": 128,
   "mla": 320,
   "vocab": 131072,
   "maxCtx": 262144,
   "native": {
    "label": "FP8 (official)",
    "gib": 111
   },
   "released": "2026-03",
   "license": "Apache 2.0",
   "hf": "mistralai/Mistral-Small-4-119B-2603",
   "notes": "Despite the name, Mistral Small 4 is a 119B mixture-of-experts (128 experts, 4 active plus a shared expert, about 6.5B active). It uses DeepSeek-style latent attention, caching a 320-value latent per layer, so its long-context cache is very small. The official release is FP8; an NVFP4 checkpoint is also available for Blackwell GPUs.",
   "faq": [
    {
     "q": "Is Mistral Small 4 actually small?",
     "a": "No. It has 119B total parameters, so even at Q4_K_M it needs about 68 GB. It is fast for its size because only about 6.5B parameters are active per token."
    }
   ]
  },
  {
   "slug": "ministral-3-3b",
   "name": "Ministral 3 3B",
   "family": "Mistral",
   "org": "Mistral AI",
   "params": 3.43,
   "layers": 26,
   "hidden": 3072,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 131072,
   "inter": 9216,
   "maxCtx": 262144,
   "released": "2025-12",
   "license": "Apache 2.0",
   "hf": "mistralai/Ministral-3-3B-Instruct-2512",
   "notes": "The smallest Ministral 3, a dense edge model with vision and a 256K context window. It uses plain grouped-query attention, so long contexts are not free: 128K of cache in FP16 is larger than the model itself. Keep the context to what you need on small GPUs.",
   "faq": [
    {
     "q": "Can Ministral 3 3B run on a 4GB GPU?",
     "a": "Yes at Q4_K_M with up to about 8K of context. Longer contexts need a Q8_0 cache or more memory."
    }
   ]
  },
  {
   "slug": "ministral-3-8b",
   "name": "Ministral 3 8B",
   "family": "Mistral",
   "org": "Mistral AI",
   "params": 8.49,
   "layers": 34,
   "hidden": 4096,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 131072,
   "inter": 14336,
   "maxCtx": 262144,
   "released": "2025-12",
   "license": "Apache 2.0",
   "hf": "mistralai/Ministral-3-8B-Instruct-2512",
   "notes": "A dense 8B model with vision and a 256K context, released alongside Mistral Large 3. It is a drop-in alternative to Llama 3.1 8B with a similar memory profile: 136 KB of cache per token in FP16. Base, instruct and reasoning variants share the same architecture.",
   "faq": [
    {
     "q": "Can Ministral 3 8B run on an 8GB GPU?",
     "a": "Yes at Q4_K_M with up to about 14K of context, similar to Llama 3.1 8B. A Q8_0 cache roughly doubles that."
    }
   ]
  },
  {
   "slug": "ministral-3-14b",
   "name": "Ministral 3 14B",
   "family": "Mistral",
   "org": "Mistral AI",
   "params": 13.5,
   "layers": 40,
   "hidden": 5120,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 131072,
   "inter": 16384,
   "maxCtx": 262144,
   "released": "2025-12",
   "license": "Apache 2.0",
   "hf": "mistralai/Ministral-3-14B-Instruct-2512",
   "notes": "The largest Ministral 3: a dense 14B model with vision, aimed at 12–16 GB GPUs. Its 40 layers with 8 KV heads cost 160 KB of cache per token in FP16, so context length is the main thing to watch on a 12 GB card.",
   "faq": [
    {
     "q": "Is 12GB enough for Ministral 3 14B?",
     "a": "Yes at Q4_K_M with an 8K–16K context. For 32K or more, use a Q8_0 cache or a 16 GB card."
    }
   ]
  },
  {
   "slug": "gemma-3-1b",
   "name": "Gemma 3 1B",
   "family": "Gemma",
   "org": "Google",
   "params": 1,
   "layers": 26,
   "hidden": 1152,
   "heads": 4,
   "kvHeads": 1,
   "headDim": 256,
   "vocab": 262144,
   "inter": 6912,
   "maxCtx": 32768,
   "swa": {
    "window": 512,
    "global": 4
   },
   "released": "2025-03",
   "license": "Gemma Terms of Use",
   "hf": "google/gemma-3-1b-it",
   "notes": "Google's smallest Gemma 3, text only. It uses a single KV head and sliding-window attention on most layers, so its context memory is tiny. The 262K vocabulary takes up about a third of its parameters.",
   "faq": [
    {
     "q": "How much RAM does Gemma 3 1B need?",
     "a": "Under 2 GB at Q4_K_M even with its full 32K context, so it runs on almost any device."
    }
   ]
  },
  {
   "slug": "gemma-3-4b",
   "name": "Gemma 3 4B",
   "family": "Gemma",
   "org": "Google",
   "params": 4.3,
   "layers": 34,
   "hidden": 2560,
   "heads": 8,
   "kvHeads": 4,
   "headDim": 256,
   "vocab": 262208,
   "inter": 10240,
   "maxCtx": 131072,
   "swa": {
    "window": 1024,
    "global": 5
   },
   "released": "2025-03",
   "license": "Gemma Terms of Use",
   "hf": "google/gemma-3-4b-it",
   "notes": "A multimodal model with a 1024-token sliding window on five of every six layers. Only five layers keep the full context, so a 128K window costs a fraction of what it would on a standard transformer. Add about 0.8 GB if you load the vision encoder.",
   "faq": [
    {
     "q": "Why does Gemma 3 use so little memory for long context?",
     "a": "Most of its layers only attend to the last 1,024 tokens, so only 5 of 34 layers store the full context in the KV cache."
    }
   ]
  },
  {
   "slug": "gemma-3-12b",
   "name": "Gemma 3 12B",
   "family": "Gemma",
   "org": "Google",
   "params": 12.2,
   "layers": 48,
   "hidden": 3840,
   "heads": 16,
   "kvHeads": 8,
   "headDim": 256,
   "vocab": 262208,
   "inter": 15360,
   "maxCtx": 131072,
   "swa": {
    "window": 1024,
    "global": 8
   },
   "released": "2025-03",
   "license": "Gemma Terms of Use",
   "hf": "google/gemma-3-12b-it",
   "notes": "A strong multimodal mid-size model. Its 256-dimension heads make each cached token heavy on the 8 full-attention layers, but sliding windows on the other 40 layers keep long contexts affordable. Google also published official QAT (quantization-aware trained) Q4_0 weights.",
   "faq": [
    {
     "q": "Can Gemma 3 12B run on 8GB VRAM?",
     "a": "At Q4_K_M the weights are about 6.9 GB, so 8 GB cards fit it only with a short context. 12 GB is the comfortable minimum."
    }
   ]
  },
  {
   "slug": "gemma-3-27b",
   "name": "Gemma 3 27B",
   "family": "Gemma",
   "org": "Google",
   "params": 27.4,
   "layers": 62,
   "hidden": 5376,
   "heads": 32,
   "kvHeads": 16,
   "headDim": 128,
   "vocab": 262208,
   "inter": 21504,
   "maxCtx": 131072,
   "swa": {
    "window": 1024,
    "global": 10
   },
   "released": "2025-03",
   "license": "Gemma Terms of Use",
   "hf": "google/gemma-3-27b-it",
   "notes": "The largest Gemma 3. It has 16 KV heads, twice as many as most models its size, but sliding windows on 52 of 62 layers offset that. Google's QAT Q4_0 release was made to fit a 24 GB consumer GPU. Older runtimes without sliding-window support store the full cache on every layer and use far more memory.",
   "faq": [
    {
     "q": "Does Gemma 3 27B fit on an RTX 3090 or 4090?",
     "a": "Yes. At Q4_K_M it needs about 18 GB with an 8K context, and the sliding window keeps 32K context under 24 GB in runtimes that support it."
    }
   ]
  },
  {
   "slug": "gemma-4-e2b",
   "name": "Gemma 4 E2B",
   "family": "Gemma",
   "org": "Google",
   "params": 5.1,
   "layers": 35,
   "hidden": 1536,
   "heads": 8,
   "kvHeads": 1,
   "headDim": 256,
   "vocab": 262144,
   "inter": 6144,
   "maxCtx": 131072,
   "attn": [
    {
     "n": 3,
     "width": 1024
    },
    {
     "n": 12,
     "width": 512,
     "window": 512
    }
   ],
   "attnNote": "Only 15 of 35 layers have their own KV cache; the last 20 reuse caches from earlier layers. Of those 15, 12 use a 512-token sliding window and 3 keep the full context (1 KV head, 512-dimension keys and values), so the cache is tiny.",
   "released": "2026-04",
   "license": "Apache 2.0",
   "hf": "google/gemma-4-E2B-it",
   "notes": "An on-device model with about 2.3B effective parameters. The file holds 5.1B parameters because each layer has its own small per-layer embedding table; runtimes can keep those tables in system memory, so real GPU use can be lower than shown here. It accepts text, images and audio.",
   "faq": [
    {
     "q": "Why is Gemma 4 E2B bigger than 2B?",
     "a": "The E stands for effective. It has 2.3B parameters in the compute path plus per-layer embedding tables, for 5.1B in total. The tables are cheap lookups, so speed is like a 2B model."
    }
   ]
  },
  {
   "slug": "gemma-4-e4b",
   "name": "Gemma 4 E4B",
   "family": "Gemma",
   "org": "Google",
   "params": 8,
   "layers": 42,
   "hidden": 2560,
   "heads": 8,
   "kvHeads": 2,
   "headDim": 256,
   "vocab": 262144,
   "inter": 10240,
   "maxCtx": 131072,
   "attn": [
    {
     "n": 4,
     "width": 2048
    },
    {
     "n": 20,
     "width": 1024,
     "window": 512
    }
   ],
   "attnNote": "Only 24 of 42 layers have their own KV cache; the last 18 reuse caches from earlier layers. Of those 24, 20 use a 512-token sliding window and 4 keep the full context (2 KV heads, 512-dimension heads), so long contexts are very cheap.",
   "released": "2026-04",
   "license": "Apache 2.0",
   "hf": "google/gemma-4-E4B-it",
   "notes": "The larger on-device Gemma 4 with about 4.5B effective parameters (8B including per-layer embeddings). It handles text, images and audio, and a full 128K context costs only about 2 GB of cache. A strong pick for 8 GB GPUs and laptops.",
   "faq": [
    {
     "q": "Can Gemma 4 E4B run on a 6GB GPU?",
     "a": "Yes at Q4_K_M. The full file is about 4.5 GB and the cache is very small; runtimes that keep per-layer embeddings in system RAM need even less VRAM."
    }
   ]
  },
  {
   "slug": "gemma-4-26b-a4b",
   "name": "Gemma 4 26B-A4B",
   "family": "Gemma",
   "org": "Google",
   "params": 25.2,
   "active": 3.8,
   "layers": 30,
   "hidden": 2816,
   "heads": 16,
   "kvHeads": 8,
   "headDim": 256,
   "vocab": 262144,
   "maxCtx": 262144,
   "attn": [
    {
     "n": 5,
     "width": 2048
    },
    {
     "n": 25,
     "width": 4096,
     "window": 1024
    }
   ],
   "attnNote": "25 of 30 layers use a 1,024-token sliding window. The 5 global layers keep the full context with only 2 KV heads of 512 dimensions, so each extra 1,000 tokens costs only about 20 MB in FP16. Figures assume separate keys and values; runtimes that store Gemma 4's unified key-value once need even less.",
   "released": "2026-04",
   "license": "Apache 2.0",
   "hf": "google/gemma-4-26B-A4B-it",
   "notes": "Google's first Gemma mixture-of-experts: 128 experts plus a shared expert, 8 routed per token, about 3.8B active. It runs at small-model speed and just squeezes onto a 16 GB card at Q4_K_M; IQ4_XS leaves more room. Long contexts are cheap thanks to sliding-window layers and very lean global layers.",
   "faq": [
    {
     "q": "Can Gemma 4 26B-A4B run on a 16GB GPU?",
     "a": "Just. At Q4_K_M it needs about 15.6 GB with an 8K context, which is tight on 16 GB; IQ4_XS is more comfortable. The small cache means a 32K context adds only about 0.5 GB."
    }
   ]
  },
  {
   "slug": "gemma-4-31b",
   "name": "Gemma 4 31B",
   "family": "Gemma",
   "org": "Google",
   "params": 30.7,
   "layers": 60,
   "hidden": 5376,
   "heads": 32,
   "kvHeads": 16,
   "headDim": 256,
   "vocab": 262144,
   "inter": 21504,
   "maxCtx": 262144,
   "attn": [
    {
     "n": 10,
     "width": 4096
    },
    {
     "n": 50,
     "width": 8192,
     "window": 1024
    }
   ],
   "attnNote": "50 of 60 layers use a 1,024-token sliding window. The 10 global layers keep the full context with 4 KV heads of 512 dimensions. Figures assume separate keys and values; runtimes that store Gemma 4's unified key-value once need about half as much on the global layers.",
   "released": "2026-04",
   "license": "Apache 2.0",
   "hf": "google/gemma-4-31B-it",
   "notes": "The largest dense Gemma 4. Its 256K context is affordable because five of every six layers only look at the last 1,024 tokens. At Q4_K_M it is a natural fit for 24 GB cards, and its vision encoder adds about 0.5 GB when loaded.",
   "faq": [
    {
     "q": "Does Gemma 4 31B fit on an RTX 3090 or 4090?",
     "a": "Yes at Q4_K_M with up to about 44K of context (about 90K with a Q8_0 cache). The weights are about 17 GB, and the cache grows slowly because only 10 of 60 layers keep the full context."
    }
   ]
  },
  {
   "slug": "phi-4-mini",
   "name": "Phi-4 Mini",
   "family": "Phi",
   "org": "Microsoft",
   "params": 3.84,
   "layers": 32,
   "hidden": 3072,
   "heads": 24,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 200064,
   "inter": 8192,
   "maxCtx": 131072,
   "released": "2025-02",
   "license": "MIT",
   "hf": "microsoft/Phi-4-mini-instruct",
   "notes": "A 3.8B model trained heavily on synthetic reasoning data, with a 200K vocabulary for better multilingual coverage. MIT-licensed, so it suits commercial edge use.",
   "faq": [
    {
     "q": "What is the smallest GPU for Phi-4 Mini?",
     "a": "About 4.1 GB at Q4_K_M with an 8K context, so 6 GB cards are comfortable. A 4 GB card works with about 5K of context."
    }
   ]
  },
  {
   "slug": "phi-4",
   "name": "Phi-4 14B",
   "family": "Phi",
   "org": "Microsoft",
   "params": 14.7,
   "layers": 40,
   "hidden": 5120,
   "heads": 40,
   "kvHeads": 10,
   "headDim": 128,
   "vocab": 100352,
   "inter": 17920,
   "maxCtx": 16384,
   "released": "2024-12",
   "license": "MIT",
   "hf": "microsoft/phi-4",
   "notes": "A 14B dense model focused on reasoning and maths. Its 16K context limit keeps total memory predictable: Q4_K_M at full context needs about 12.2 GB, or 10.7 GB with a Q8_0 KV cache.",
   "faq": [
    {
     "q": "Does Phi-4 fit on a 12GB GPU?",
     "a": "Yes with a little tuning. At Q4_K_M it needs about 12.2 GB at its full 16K context, so on 12 GB use a Q8_0 KV cache (about 10.7 GB) or keep the context near 8K."
    }
   ]
  },
  {
   "slug": "deepseek-r1",
   "name": "DeepSeek R1 671B",
   "family": "DeepSeek",
   "org": "DeepSeek",
   "params": 671,
   "active": 37,
   "layers": 61,
   "hidden": 7168,
   "heads": 128,
   "kvHeads": 128,
   "headDim": 128,
   "mla": 576,
   "vocab": 129280,
   "maxCtx": 131072,
   "native": {
    "label": "FP8 (official)",
    "gib": 641
   },
   "released": "2025-01",
   "license": "MIT",
   "hf": "deepseek-ai/DeepSeek-R1",
   "notes": "The full reasoning model, a 671B mixture-of-experts with 37B active parameters. It uses Multi-head Latent Attention, which compresses the KV cache to 576 values per token per layer, about 69 KB per token in FP16. Runtimes without MLA support store full keys and values and need many times more memory for context. The official checkpoint is FP8, about 641 GiB.",
   "faq": [
    {
     "q": "Can I run the full DeepSeek R1 at home?",
     "a": "Only with very large unified memory or a server: about 380 GB at Q4_K_M. Community dynamic 1.58-bit and 2-bit quants bring it to roughly 130 to 230 GB at a quality cost."
    },
    {
     "q": "Are the DeepSeek R1 distill models the same as R1?",
     "a": "No. The distills are Qwen and Llama models fine-tuned on R1's outputs. They are far smaller and have their own pages here."
    }
   ]
  },
  {
   "slug": "deepseek-r1-distill-qwen-7b",
   "name": "DeepSeek R1 Distill Qwen 7B",
   "family": "DeepSeek",
   "org": "DeepSeek",
   "params": 7.62,
   "layers": 28,
   "hidden": 3584,
   "heads": 28,
   "kvHeads": 4,
   "headDim": 128,
   "vocab": 152064,
   "inter": 18944,
   "maxCtx": 131072,
   "released": "2025-01",
   "license": "MIT",
   "hf": "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
   "notes": "Qwen 2.5 Math 7B fine-tuned on DeepSeek R1 reasoning traces. Memory needs match Qwen 2.5 7B, but reasoning traces are long, so budget for 16K to 32K of context.",
   "faq": [
    {
     "q": "How much VRAM does R1 Distill Qwen 7B need for reasoning?",
     "a": "About 6.9 GB at Q4_K_M with 32K of context, thanks to its small 4-head KV cache. An 8 GB card handles it."
    }
   ]
  },
  {
   "slug": "deepseek-r1-distill-llama-8b",
   "name": "DeepSeek R1 Distill Llama 8B",
   "family": "DeepSeek",
   "org": "DeepSeek",
   "params": 8.03,
   "layers": 32,
   "hidden": 4096,
   "heads": 32,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 128256,
   "inter": 14336,
   "maxCtx": 131072,
   "released": "2025-01",
   "license": "MIT (Llama 3.1 license applies)",
   "hf": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
   "notes": "Llama 3.1 8B fine-tuned on R1 reasoning traces. Same memory profile as Llama 3.1 8B, with a KV cache more than twice the size of the Qwen 7B distill's per token.",
   "faq": [
    {
     "q": "Llama 8B or Qwen 7B distill for an 8GB GPU?",
     "a": "The Qwen 7B distill fits longer reasoning traces on 8 GB because its KV cache is less than half the size per token."
    }
   ]
  },
  {
   "slug": "deepseek-r1-distill-qwen-14b",
   "name": "DeepSeek R1 Distill Qwen 14B",
   "family": "DeepSeek",
   "org": "DeepSeek",
   "params": 14.7,
   "layers": 48,
   "hidden": 5120,
   "heads": 40,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 152064,
   "inter": 13824,
   "maxCtx": 131072,
   "released": "2025-01",
   "license": "MIT",
   "hf": "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B",
   "notes": "Qwen 2.5 14B fine-tuned on R1 traces. The 48 layers make long reasoning expensive: 32K of context adds 6 GB in FP16. A Q8_0 KV cache is the easiest win on 12 GB and 16 GB cards.",
   "faq": [
    {
     "q": "Does R1 Distill Qwen 14B fit on a 16GB GPU?",
     "a": "Yes at Q4_K_M with up to about 30K of FP16 context, or more with a quantized KV cache."
    }
   ]
  },
  {
   "slug": "deepseek-r1-distill-qwen-32b",
   "name": "DeepSeek R1 Distill Qwen 32B",
   "family": "DeepSeek",
   "org": "DeepSeek",
   "params": 32.8,
   "layers": 64,
   "hidden": 5120,
   "heads": 40,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 152064,
   "inter": 27648,
   "maxCtx": 131072,
   "released": "2025-01",
   "license": "MIT",
   "hf": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
   "notes": "The strongest of the small distills and a favourite for 24 GB cards. Reasoning traces fill context quickly, and each 1K tokens costs 256 MB in FP16, so a Q8_0 cache is worth turning on.",
   "faq": [
    {
     "q": "How long a reasoning trace fits on 24GB with R1 Distill 32B?",
     "a": "At Q4_K_M, about 12K tokens with an FP16 cache, or about 23K with a Q8_0 cache."
    }
   ]
  },
  {
   "slug": "deepseek-r1-distill-llama-70b",
   "name": "DeepSeek R1 Distill Llama 70B",
   "family": "DeepSeek",
   "org": "DeepSeek",
   "params": 70.6,
   "layers": 80,
   "hidden": 8192,
   "heads": 64,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 128256,
   "inter": 28672,
   "maxCtx": 131072,
   "released": "2025-01",
   "license": "MIT (Llama 3.3 license applies)",
   "hf": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
   "notes": "Llama 3.3 70B fine-tuned on R1 reasoning traces. Memory needs are identical to Llama 3.3 70B, but long reasoning means planning for 16K to 32K of context, which adds 5 to 10 GB in FP16.",
   "faq": [
    {
     "q": "What do I need to run R1 Distill Llama 70B?",
     "a": "48 GB of VRAM (two 24 GB cards) for Q4_K_M with about 12K to 14K of context, or 64 GB+ of unified memory on a Mac or Ryzen AI Max."
    }
   ]
  },
  {
   "slug": "deepseek-v4-flash",
   "name": "DeepSeek V4 Flash",
   "family": "DeepSeek",
   "org": "DeepSeek",
   "params": 284,
   "active": 13,
   "layers": 43,
   "hidden": 4096,
   "heads": 64,
   "kvHeads": 1,
   "headDim": 512,
   "vocab": 129280,
   "maxCtx": 262144,
   "attn": [
    {
     "n": 43,
     "width": 512,
     "window": 128
    },
    {
     "n": 20,
     "width": 640,
     "ratio": 4
    },
    {
     "n": 20,
     "width": 512,
     "ratio": 128
    }
   ],
   "attnNote": "Compressed attention: every layer keeps a 128-token local window, 20 layers also keep one compressed entry per 4 tokens (plus indexer keys) and 20 keep one per 128 tokens. That makes a long context very cheap. These cache figures are our estimate from config.json; real use depends on runtime support.",
   "native": {
    "label": "FP4 + FP8 (official)",
    "gib": 143
   },
   "released": "2026-04",
   "license": "MIT",
   "hf": "deepseek-ai/DeepSeek-V4-Flash",
   "notes": "The smaller DeepSeek V4: 284B total parameters with about 13B active and a 1M-token context. Experts ship in FP4 and most other weights in FP8, so the official checkpoint is about 143 GB by our estimate. Its compressed attention keeps the cache a small fraction of earlier DeepSeek models. We cap context at 256K here.",
   "faq": [
    {
     "q": "Can DeepSeek V4 Flash run on a 192GB or 256GB Mac?",
     "a": "Yes. The native FP4 + FP8 weights are about 143 GB by our estimate, and the compressed cache is small, so a 192 GB machine with a raised GPU memory limit or a 256 GB Mac can hold it."
    }
   ]
  },
  {
   "slug": "gpt-oss-20b",
   "name": "gpt-oss-20b",
   "family": "gpt-oss",
   "org": "OpenAI",
   "params": 20.9,
   "active": 3.6,
   "layers": 24,
   "hidden": 2880,
   "heads": 64,
   "kvHeads": 8,
   "headDim": 64,
   "vocab": 201088,
   "maxCtx": 131072,
   "swa": {
    "window": 128,
    "global": 12
   },
   "native": {
    "label": "MXFP4 (official)",
    "gib": 12.8
   },
   "released": "2025-08",
   "license": "Apache 2.0",
   "hf": "openai/gpt-oss-20b",
   "notes": "OpenAI's smaller open-weight model, a mixture-of-experts with 3.6B active parameters. It ships natively in MXFP4, so the official checkpoint is already about 13 GB and re-quantizing to GGUF saves little. Half the layers use a 128-token sliding window, which keeps long contexts cheap.",
   "faq": [
    {
     "q": "Can gpt-oss-20b run on a 16GB GPU?",
     "a": "Yes. The native MXFP4 weights are about 12.8 GiB, leaving room for a long context on 16 GB cards. OpenAI designed it for 16 GB devices."
    },
    {
     "q": "Should I use a Q4_K_M GGUF of gpt-oss-20b?",
     "a": "There is little benefit. The experts are already 4-bit, so most GGUF versions keep MXFP4 for them and end up close to the native size."
    }
   ]
  },
  {
   "slug": "gpt-oss-120b",
   "name": "gpt-oss-120b",
   "family": "gpt-oss",
   "org": "OpenAI",
   "params": 116.8,
   "active": 5.1,
   "layers": 36,
   "hidden": 2880,
   "heads": 64,
   "kvHeads": 8,
   "headDim": 64,
   "vocab": 201088,
   "maxCtx": 131072,
   "swa": {
    "window": 128,
    "global": 18
   },
   "native": {
    "label": "MXFP4 (official)",
    "gib": 60.8
   },
   "released": "2025-08",
   "license": "Apache 2.0",
   "hf": "openai/gpt-oss-120b",
   "notes": "OpenAI's larger open-weight model: 117B total and 5.1B active parameters, shipped natively in MXFP4 at about 61 GiB. It fits on a single 80 GB H100 or a 96 GB RTX PRO 6000, and it is fast on 128 GB unified-memory systems because so little is read per token.",
   "faq": [
    {
     "q": "Can gpt-oss-120b run on a 128GB Mac or Ryzen AI Max?",
     "a": "Yes. About 61 GiB of weights plus cache fits in the GPU-usable portion of 128 GB, and only 5.1B active parameters keep it fast."
    },
    {
     "q": "Can gpt-oss-120b run on a 24GB GPU?",
     "a": "Only with expert offload to system RAM (64 GB+). Attention layers stay on the GPU and the experts run on the CPU, which still gives usable speed."
    }
   ]
  },
  {
   "slug": "glm-4-5-air",
   "name": "GLM-4.5-Air",
   "family": "GLM",
   "org": "Z.ai",
   "params": 106,
   "active": 12,
   "layers": 46,
   "hidden": 4096,
   "heads": 96,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 151552,
   "maxCtx": 131072,
   "released": "2025-07",
   "license": "MIT",
   "hf": "zai-org/GLM-4.5-Air",
   "notes": "A 106B mixture-of-experts with 12B active parameters that became a favourite for 96–128 GB machines. It uses standard grouped-query attention across 46 layers, so a long context adds memory steadily: 184 KB per token in FP16. Strong at coding and tool use for its active size.",
   "faq": [
    {
     "q": "Can GLM-4.5-Air run on 64GB of RAM and a 24GB GPU?",
     "a": "Yes with partial offload. Keep attention and shared layers on the GPU and the experts in system RAM; with only 12B active parameters it still reaches usable speed."
    }
   ]
  },
  {
   "slug": "minimax-m2",
   "name": "MiniMax-M2",
   "family": "MiniMax",
   "org": "MiniMax",
   "params": 230,
   "active": 10,
   "layers": 62,
   "hidden": 3072,
   "heads": 48,
   "kvHeads": 8,
   "headDim": 128,
   "vocab": 200064,
   "maxCtx": 196608,
   "native": {
    "label": "FP8 (official)",
    "gib": 215
   },
   "released": "2025-10",
   "license": "MIT",
   "hf": "MiniMaxAI/MiniMax-M2",
   "notes": "A 230B mixture-of-experts with only 10B active parameters, built for coding agents. The official weights are FP8. Every one of its 62 layers uses full attention, so the KV cache grows quickly with long agent sessions: about 248 KB per token in FP16.",
   "faq": [
    {
     "q": "What do I need to run MiniMax-M2 locally?",
     "a": "At Q4_K_M about 130 GB plus cache, so a 192 GB or 512 GB Mac Studio, or four to six 24–48 GB GPUs. A 128 GB machine needs a Q2-level quant."
    }
   ]
  },
  {
   "slug": "kimi-k2-6",
   "name": "Kimi K2.6",
   "family": "Kimi",
   "org": "Moonshot AI",
   "params": 1000,
   "active": 32,
   "layers": 61,
   "hidden": 7168,
   "heads": 64,
   "kvHeads": 64,
   "headDim": 128,
   "mla": 576,
   "vocab": 163840,
   "maxCtx": 262144,
   "native": {
    "label": "INT4 (official)",
    "gib": 555
   },
   "released": "2026-04",
   "license": "Modified MIT",
   "hf": "moonshotai/Kimi-K2.6",
   "notes": "A one-trillion-parameter MoE (384 experts, 8 active, about 32B active) on the DeepSeek V3 architecture with latent attention. The experts ship natively in 4-bit integer format, so the official checkpoint is roughly 555 GB and further quantization saves little. The native INT4 size here is our estimate from the quantization config.",
   "faq": [
    {
     "q": "Can Kimi K2.6 run on a 512GB Mac Studio?",
     "a": "Only with lower-bit quants. The native INT4 weights are about 555 GB, more than the roughly 384 GB the GPU can use by default, so you need an IQ2 or IQ3 GGUF or a raised wired-memory limit."
    }
   ]
  }
 ]
}