{  "version": 1,  "updated": "2026-08-18",  "tiers": [    {      "id": "nano",      "label": "Nano",      "blurb": "Single consumer GPU / laptop",      "models": [        {          "id": "qwen3.5-9b",          "name": "Qwen3.5-9B",          "role": null,          "description": "Dense 9B with vision encoder, hybrid Gated DeltaNet/attention, 262K context",          "params": "9B dense, hybrid Gated DeltaNet/attention",          "active_params_b": null,          "context_tokens": 262144,          "license": "Apache-2.0",          "multimodal": true,          "notes": [            "Listed as \"qwen3.6-9b\" in some sources — no 3.6 9B exists; the 9B dense is Qwen3.5",            "Thinks by default — serve with --reasoning-parser qwen3",            "Needs bleeding-edge runtimes (vLLM nightly)",            "Text-only serving: --language-model-only frees vision memory",            "GGUF needs mmproj-*.gguf for vision",            "No official FP8 or NVFP4 — RedHatAI (FP8) and community (NVFP4) only"          ],          "variants": [            {              "format": "bf16",              "repo": "Qwen/Qwen3.5-9B",              "official": true,              "size_gb": 19.3,              "caveat": null            },            {              "format": "fp8",              "repo": "RedHatAI/Qwen3.5-9B-FP8-dynamic",              "official": false,              "source": "RedHatAI",              "size_gb": null,              "caveat": null            },            {              "format": "nvfp4",              "repo": "kaitchup/Qwen3.5-9B-autoround-NVFP4",              "official": false,              "source": "kaitchup",              "size_gb": null,              "caveat": "community quant"            },            {              "format": "q4",              "repo": "unsloth/Qwen3.5-9B-GGUF",              "official": false,              "source": "unsloth",              "allow_patterns": [                "*Q4_K_M*.gguf"              ],              "size_gb": null,              "caveat": null            }          ],          "architecture": "Dense · hybrid Gated DeltaNet + attention",          "total_params_b": 9.0,          "intelligence_index": 21.8,          "agentic_index": 7.0        },        {          "id": "gemma-4-e2b",          "name": "Gemma 4 E2B",          "role": null,          "description": "2.3B effective multimodal (text+image+audio in) with 128K context",          "params": "5.1B total w/ Per-Layer Embeddings (2.3B effective)",          "active_params_b": null,          "context_tokens": 131072,          "license": "Apache-2.0",          "multimodal": true,          "notes": [            "NVFP4 needs vLLM ≥ 0.25 + flashinfer — let vLLM auto-select the kernel (not Marlin)",            "Google also ships official QAT 4-bit GGUFs (google/gemma-4-E2B-it-qat-q4_0-gguf) as a creator-official alternative",            "No official FP8 — community FP8-dynamic only; official 4-bit path is Google's own QAT GGUFs"          ],          "variants": [            {              "format": "bf16",              "repo": "google/gemma-4-E2B-it",              "official": true,              "size_gb": null,              "caveat": null            },            {              "format": "fp8",              "repo": "leon-se/gemma-4-E2B-it-FP8-Dynamic",              "official": false,              "source": "leon-se",              "size_gb": null,              "caveat": "community quant"            },            {              "format": "nvfp4",              "repo": "unsloth/gemma-4-E2B-it-NVFP4",              "official": false,              "source": "unsloth",              "size_gb": null,              "caveat": null            },            {              "format": "q4",              "repo": "unsloth/gemma-4-E2B-it-GGUF",              "official": false,              "source": "unsloth",              "allow_patterns": [                "*Q4_K_M*.gguf"              ],              "size_gb": null,              "caveat": null            },            {              "format": "q4",              "repo": "google/gemma-4-E2B-it-qat-q4_0-gguf",              "official": true,              "size_gb": null,              "caveat": null            }          ],          "architecture": "Dense · Per-Layer Embeddings",          "total_params_b": 5.1,          "intelligence_index": 9.5,          "agentic_index": null        },        {          "id": "lfm2.5-2.6b",          "name": "LFM2.5-2.6B",          "role": "fast",          "description": "Liquid's 2.6B hybrid built for on-device agents — 128K context, tool use, and RAG",          "params": "2.69B dense, LFM2 hybrid conv/attention",          "architecture": "Dense · LFM2 hybrid (short conv + GQA)",          "total_params_b": 2.69,          "active_params_b": null,          "context_tokens": 131072,          "license": "LFM Open License v1.0",          "multimodal": false,          "notes": [            "Built for on-device agentic workloads — tool use, extraction, RAG; not a coding or knowledge model",            "Pre-trained on ~34T tokens, mid-trained to a 128K window",            "Ships official GGUF, MLX, and ONNX builds alongside the safetensors checkpoint",            "No official FP8 or NVFP4 — it is small enough that BF16 fits almost anywhere"          ],          "variants": [            {              "format": "bf16",              "repo": "LiquidAI/LFM2.5-2.6B",              "official": true,              "size_gb": 5.4,              "caveat": null            },            {              "format": "q4",              "repo": "LiquidAI/LFM2.5-2.6B-GGUF",              "official": true,              "allow_patterns": [                "*Q4_K_M*.gguf"              ],              "size_gb": 1.7,              "caveat": null            }          ],          "intelligence_index": null,          "agentic_index": null        }      ]    },    {      "id": "mini",      "label": "Mini",      "blurb": "Single 24–48 GB GPU",      "models": [        {          "id": "gemma-4-26b-a4b",          "name": "Gemma 4 26B A4B",          "role": "fast",          "description": "Multimodal MoE with only 3.8B active params — 3–5× faster decode than dense picks",          "params": "25.2B total MoE",          "active_params_b": 3.8,          "context_tokens": 262144,          "license": "Gemma",          "multimodal": true,          "notes": [            "MoE ≠ small download — all weights must still fit in memory",            "Do NOT use vLLM --quantization fp8 on-the-fly with the BF16 repo (broken output — use the RedHatAI checkpoint)",            "NVFP4 on non-Blackwell falls back to Marlin and is slower than FP8",            "GGUF has a known ROCm infinite-loop bug; bartowski publishes alternative imatrix GGUFs"          ],          "variants": [            {              "format": "bf16",              "repo": "google/gemma-4-26B-A4B-it",              "official": true,              "size_gb": 52,              "caveat": null            },            {              "format": "fp8",              "repo": "RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic",              "official": false,              "source": "RedHatAI",              "size_gb": null,              "caveat": null            },            {              "format": "nvfp4",              "repo": "nvidia/Gemma-4-26B-A4B-NVFP4",              "official": false,              "source": "nvidia",              "size_gb": null,              "caveat": null            },            {              "format": "q4",              "repo": "unsloth/gemma-4-26B-A4B-it-GGUF",              "official": false,              "source": "unsloth",              "allow_patterns": [                "*UD-Q4_K_M*.gguf"              ],              "size_gb": null,              "caveat": null            }          ],          "architecture": "MoE · multimodal",          "total_params_b": 25.2,          "intelligence_index": 26.1,          "agentic_index": 11.0        },        {          "id": "gemma-4-31b",          "name": "Gemma 4 31B",          "role": "smart",          "description": "Dense 31B with 262K context — the higher-quality dense pick in this tier",          "params": "31B dense",          "active_params_b": null,          "context_tokens": 262144,          "license": "Gemma",          "multimodal": false,          "notes": [            "FP8-block (W8A8, better accuracy) or FP8-dynamic, both need sm_89+",            "NVFP4 is Blackwell-only",            "Unsloth repo includes MTP draft GGUFs for speculative decoding"          ],          "variants": [            {              "format": "bf16",              "repo": "google/gemma-4-31B-it",              "official": true,              "size_gb": 62,              "caveat": null            },            {              "format": "fp8",              "repo": "RedHatAI/gemma-4-31B-it-FP8-block",              "official": false,              "source": "RedHatAI",              "size_gb": null,              "caveat": null            },            {              "format": "nvfp4",              "repo": "nvidia/Gemma-4-31B-IT-NVFP4",              "official": false,              "source": "nvidia",              "size_gb": null,              "caveat": null            },            {              "format": "q4",              "repo": "unsloth/gemma-4-31B-it-GGUF",              "official": false,              "source": "unsloth",              "allow_patterns": [                "*Q4_K_M*.gguf"              ],              "size_gb": null,              "caveat": null            }          ],          "architecture": "Dense",          "total_params_b": 31.0,          "intelligence_index": 29.7,          "agentic_index": 14.4        },        {          "id": "qwen3.6-35b-a3b",          "name": "Qwen3.6-35B-A3B",          "role": "fast",          "description": "256-expert MoE with only 3B active params and MTP speculative decoding built in",          "params": "35B MoE, 256 experts",          "active_params_b": 3,          "context_tokens": 262144,          "license": "Apache-2.0",          "multimodal": false,          "notes": [            "MTP built in (usable as speculative decoding)",            "Needs vLLM ≥ 0.19 / SGLang ≥ 0.5.10",            "GGUF arch qwen35moe — older llama.cpp won't load it",            "Q4_K_M ≈ 20 GB"          ],          "variants": [            {              "format": "bf16",              "repo": "Qwen/Qwen3.6-35B-A3B",              "official": true,              "size_gb": null,              "caveat": null            },            {              "format": "fp8",              "repo": "Qwen/Qwen3.6-35B-A3B-FP8",              "official": true,              "size_gb": null,              "caveat": null            },            {              "format": "nvfp4",              "repo": "unsloth/Qwen3.6-35B-A3B-NVFP4",              "official": false,              "source": "unsloth",              "size_gb": null,              "caveat": null            },            {              "format": "q4",              "repo": "unsloth/Qwen3.6-35B-A3B-GGUF",              "official": false,              "source": "unsloth",              "allow_patterns": [                "*UD-Q4_K_M*.gguf"              ],              "size_gb": 20,              "caveat": null            }          ],          "architecture": "MoE · 256 experts + MTP",          "total_params_b": 35.0,          "intelligence_index": 32.1,          "agentic_index": 21.6        },        {          "id": "qwen3.8-27b",          "name": "Qwen3.8-27B",          "role": "smart",          "description": "Dense 27B vision-language model with 262K context — the strongest model that still fits one 24 GB GPU at Q4",          "params": "27.8B dense + vision encoder",          "architecture": "Dense · Gated DeltaNet + Gated Attention, native VLM",          "total_params_b": 27.8,          "active_params_b": null,          "context_tokens": 262144,          "license": "Apache-2.0",          "multimodal": true,          "notes": [            "Native vision-language: images and hour-scale video, not just text",            "Thinking is on by default; disable per request and tune depth with reasoning_effort",            "Text-only serving: --language-model-only frees the vision encoder's memory",            "Q4_K_M ≈ 17 GB, so a 24 GB GPU is the realistic floor",            "Official BF16 and FP8 from Qwen; NVFP4 and AWQ are community builds"          ],          "variants": [            {              "format": "bf16",              "repo": "Qwen/Qwen3.8-27B",              "official": true,              "size_gb": 55.6,              "caveat": null            },            {              "format": "fp8",              "repo": "Qwen/Qwen3.8-27B-FP8",              "official": true,              "size_gb": 30.9,              "caveat": null            },            {              "format": "nvfp4",              "repo": "unsloth/Qwen3.8-27B-NVFP4",              "official": false,              "source": "unsloth",              "size_gb": 23.4,              "caveat": null            },            {              "format": "q4",              "repo": "unsloth/Qwen3.8-27B-GGUF",              "official": false,              "source": "unsloth",              "allow_patterns": [                "*Q4_K_M*.gguf"              ],              "size_gb": 17.1,              "caveat": null            },            {              "format": "q4",              "repo": "cyankiwi/Qwen3.8-27B-AWQ-INT4",              "official": false,              "source": "cyankiwi",              "size_gb": 21.0,              "caveat": "AWQ INT4 for vLLM, not GGUF"            }          ],          "intelligence_index": 52.0,          "agentic_index": 50.9        }      ]    },    {      "id": "medium",      "label": "Medium",      "blurb": "Large unified memory / multi-GPU — 200B+ MoEs",      "models": [        {          "id": "step-3.7-flash",          "name": "Step 3.7 Flash",          "role": null,          "description": "198B MoE with 11B active — runs on a 128 GB Mac Studio or DGX Spark at Q4",          "params": "198B MoE",          "active_params_b": 11,          "context_tokens": 262144,          "license": "Apache-2.0",          "multimodal": false,          "notes": [            "Even Q4 needs ≥120 GB unified memory (Mac Studio 128 GB, DGX Spark)",            "llama.cpp requires StepFun's fork (branch step3.7)",            "vLLM needs the dedicated vllm/vllm-openai:stepfun37 image + --trust-remote-code --disable-cascade-attn --reasoning-parser step3p5",            "Rare: StepFun ships its own NVFP4"          ],          "variants": [            {              "format": "bf16",              "repo": "stepfun-ai/Step-3.7-Flash",              "official": true,              "size_gb": null,              "caveat": null            },            {              "format": "fp8",              "repo": "stepfun-ai/Step-3.7-Flash-FP8",              "official": true,              "size_gb": null,              "caveat": null            },            {              "format": "nvfp4",              "repo": "stepfun-ai/Step-3.7-Flash-NVFP4",              "official": true,              "size_gb": null,              "caveat": null            },            {              "format": "q4",              "repo": "unsloth/Step-3.7-Flash-GGUF",              "official": false,              "source": "unsloth",              "allow_patterns": [                "*UD-Q4_K_XL*.gguf"              ],              "size_gb": null,              "caveat": null            },            {              "format": "q4",              "repo": "stepfun-ai/Step-3.7-Flash-GGUF",              "official": true,              "allow_patterns": [                "*Q4_K_S*.gguf"              ],              "size_gb": null,              "caveat": null            }          ],          "architecture": "MoE",          "total_params_b": 198.0,          "intelligence_index": 30.9,          "agentic_index": 21.7        },        {          "id": "deepseek-v4-flash-0731",          "name": "DeepSeek V4 Flash 0731",          "role": null,          "description": "304B MoE with 1M context and a built-in speculative decoding module — the official release, superseding the preview",          "params": "304B MoE, 256 routed experts (top-6)",          "architecture": "MoE · 256 experts, top-6, + speculative decoder",          "total_params_b": 304.0,          "active_params_b": 13.0,          "context_tokens": 1048576,          "license": "MIT",          "multimodal": false,          "notes": [            "Official 0731 release — supersedes the DeepSeek-V4-Flash preview",            "Ships with a speculative decoding module attached (same structure as the DSpark build)",            "No BF16 checkpoint exists — natively FP4-experts + FP8 mixed by design",

Lines 1–500 of 723. Load more