Microsoft
BitNet b1.58
微软带来的研究范式革命。BitNet b1.58用三元{-1, 0, 1}权重替代16位权重，将矩阵乘法（MatMul）转变为简单加法，在CPU上实现极速推理与超高效率。
最低~2.8 GB （Q4_K_M）
开源大模型库
我们追踪的每一个开放权重模型系列：各个尺寸、Q4_K_M 下的显存需求，以及每个变体的独立页面。从你的显存出发，而不是从排行榜出发。
作者： Jakub Rusinowski · 最后更新：
123 个系列 · 245 个变体 · 55 家提供商按我的硬件排序
你的内存能装下什么
每档显存或统一内存能装下的最强模型，并注明计算尺寸时所用的版本。点击每一行可查看该档位的完整列表。
| 内存 | 首选 | 全部可运行 |
|---|---|---|
| 8 GB |
| 71 个模型 → |
| 12 GB |
| 85 个模型 → |
| 16 GB |
| 95 个模型 → |
| 24 GB |
| 131 个模型 → |
| 32 GB |
| 132 个模型 → |
| 48 GB |
| 138 个模型 → |
DeepSeek
DeepSeek R1
开源世界的最新轰动之作。DeepSeek R1采用混合专家架构（MoE），在推理和编程任务上媲美顶级专有模型。以出众的逻辑和数学能力著称。
Meta
Llama 3.3
Meta最新旗舰模型。Llama 3.3 70B在众多基准测试中媲美庞大的405B模型，同时效率足够高，适合高端本地部署。
最低~43.1 GB （Q4_K_M）
Meta
Llama 3.1 Family
消费级硬件开源权重的领军者。Llama 3.1拥有高达128k的上下文窗口，在众多基准测试中媲美GPT-4o。
最低~5.6 GB （Q4_K_M）
NVIDIA
Nemotron 70B
NVIDIA对Llama 3.1 70B的专业微调版本。在帮助性和指令遵循排行榜上名列前茅，对齐能力大幅超越基础Llama模型。
最低~43.4 GB （Q4_K_M）
Cohere
Command R Family
专门针对RAG（检索增强生成）和工具使用优化的模型。Command R系列在引用来源和文档推理方面表现卓越。
Microsoft
Phi-4 Family
微软最新专注于高质量合成数据训练的模型。Phi-4 14B在推理基准测试中超越许多两倍大小的模型。
最低~9.3 GB （Q4_K_M）
Microsoft
Phi 3.5 Family
微软高效的小语言模型（SLM）系列。完美适用于移动设备、边缘计算和低VRAM环境，同时保持基本推理能力。
最低~3.1 GB （Q4_K_M）
Alibaba Cloud
Qwen 2.5 Family
开源权重世界的编程与数学强者。Qwen 2.5系列在同等规模的STEM任务中始终胜过Llama 3.1。
Google
Gemma 2 Family
基于与Google Gemini相同的研究和技术构建。Gemma 2利用知识蒸馏实现极高效率和创意性能。
最低~6.2 GB （Q4_K_M）
Mistral AI
Mistral Family
欧洲开源AI的领军者。以高效率和与NVIDIA合作的Nemo模型闻名。Mistral模型是生产就绪的主力工作模型。
01.AI
Yi 1.5 Family
01.AI的高性能模型系列。以强大的英汉双语能力和编程性能著称。
IBM
Granite 3.0
IBM企业级模型系列。Granite 3.0采用Apache 2.0许可，针对编程、透明度和企业安全性进行了严格训练。
最低~5.6 GB （Q4_K_M）
Meta
Llama 4
Meta突破性的混合专家（MoE）系列。Llama 4采用稀疏MoE架构，每个token仅激活部分参数，以远低于总参数量的硬件需求提供前沿级别的智能。
最低~66.6 GB （Q4_K_M）
Google
Gemma 3
Google最强大的开源模型系列。Gemma 3提供出色的VRAM性价比，4B模型超越许多7B竞品。具备多模态能力，所有规格均支持128k上下文窗口。
Google
Gemma 4
Google第四代开源模型系列，2026年初发布。Gemma 4在保持友好硬件要求的同时，大幅提升推理、编程和多语言能力。1B和4B变体可在手机和入门GPU上运行；27B与70B级模型一争高下。
最低~3.9 GB （Q4_K_M）
Meta
Llama 3.2 Family
Meta的多模态和边缘优化系列。Llama 3.2引入视觉能力（11B和90B变体），以及专为智能手机和边缘设备设计的超紧凑1B/3B版本。所有变体均支持128k上下文窗口。
DeepSeek
DeepSeek V3
DeepSeek突破性的非推理前沿模型。仅花费550万美元训练——是同类专有模型成本的十分之一。采用685B MoE架构（每token激活37B参数），在编程、数学和通用指令遵循方面位居开源排行榜榜首。
最低~414.4 GB （Q4_K_M）
Alibaba Cloud
Qwen 3
阿里巴巴的下一代模型系列，具备突破性的混合思维模式。Qwen 3模型可按需切换快速响应模式和深度思维链推理——输入/think处理复杂问题，/no_think获得快速答案。提供密集型和MoE变体。
最低~5.8 GB （Q4_K_M）
Alibaba Cloud
Qwen 2.5 VL
阿里巴巴强大的视觉语言模型，能够理解图像、文档、图表和视频。Qwen 2.5 VL 72B在多项视觉基准测试上媲美GPT-4V，而7B版本在6GB VRAM以下即可提供强大的OCR和文档解析能力。
Mistral AI
Mistral Small 3.1
Mistral最新的小型视觉模型，Apache 2.0许可证。Mistral Small 3.1 24B具备128k上下文窗口（是前代的4倍）、多模态图像理解能力，在大多数基准测试中超越Gemma 3 27B，且完全商业可用。
最低~15 GB （Q4_K_M）
Mistral AI
Codestral
Mistral专用的代码生成模型，在覆盖80+编程语言的庞大编程语料库上训练。擅长填充中间（FIM）补全——这是IDE自动补全的核心技术。原生集成VS Code的continue.dev和Cursor。
最低~14.2 GB （Q4_K_M）
Microsoft
Phi-4 Mini
微软针对移动和边缘部署的超高效小型模型。Phi-4 Mini在仅3.8B参数内实现了卓越的推理能力，专为无独立GPU的手机和笔记本上的设备端AI设计。
最低~3.1 GB （Q4_K_M）
BigCode
StarCoder 2
顶级开源代码生成模型。StarCoder 2在涵盖600+编程语言的The Stack v2数据集上训练。专为代码补全、生成和理解构建——而非通用模型的代码微调版本。
Allen AI
OLMo 2
世界上最透明的LLM。OLMo 2公开一切：权重、训练代码、训练数据、评估套件和中间检查点。非常适合需要完全可审计AI的研究人员和合规敏感部署。在大多数基准测试中与Llama 3.1持平。
Hugging Face
SmolLM2
最小的生产质量LLM。HuggingFace的SmolLM2模型专为微控制器、手机和通过WebAssembly的浏览器设计。尽管体积极小，通过精心的数据整理展现出令人惊讶的智能。
TII (UAE)
Falcon 3
Technology Innovation Institute在14万亿token上训练的最新一代模型。Falcon 3在同等参数量上达到SOTA水平，采用完全宽松的Apache 2.0许可证，是商业部署的最佳选择之一。
Shanghai AI Lab
InternLM 3
来自中国上海AI实验室的强大开源模型。InternLM 3在中英双语任务、编程和长文档分析方面表现出色。在数学和STEM推理方面尤为突出，是中文语言应用的顶级开源选择。
Cohere
Aya Expanse
世界上最好的开源多语言模型。Aya Expanse支持23种语言，专门针对非英语用户优化。在阿拉伯语、中文、印地语、法语、西班牙语、俄语及其他17种语言上超越Llama 3.1。
Alibaba Cloud
Qwen3-Coder
阿里巴巴2026年2月发布的专用编程模型。80B-A3B变体采用稀疏MoE设计（仅3B活跃参数），超越DeepSeek V3.2，在SWE-bench Pro上媲美Claude Sonnet 4.5。在仓库级编程、调试和代码审查方面最为出色。
最低~5.6 GB （Q4_K_M）
Zhipu AI (Z.ai)
GLM-4.7 / GLM-Z1
智谱AI 2026年2月的MoE模型。GLM-4.7采用32B活跃MoE架构，在幻觉排行榜上与Claude Opus 4.5持平。GLM-Z1推理变体在人工智能指数上接近GPT-5.2水平。强大的多语言支持，尤其在中文方面有深度优势。
最低~6.2 GB （Q4_K_M）
Cohere
Aya 3B (Tiny Aya)
Cohere 2026年2月17日发布——3B参数级别最强的多语言开源模型。Tiny Aya在23种语言的多语言基准测试中超越Qwen3-4B、Gemma 3 4B和Ministral 3B。非常适合边缘设备、手机和资源受限但需要全球语言支持的部署场景。
最低~2.6 GB （Q4_K_M）
LG AI Research
EXAONE 3.5
LG AI Research的旗舰开源模型系列，在20项基准测试中获得顶级实际可用性评分。在英韩双语的指令遵循和长上下文任务中表现卓越。MIT许可证，商业可用。
Google DeepMind
Gemma 3n
Google移动优先的多模态模型系列。采用MatFormer嵌套架构——单个模型文件包含多个可按不同规格运行的子模型。处理文本、图像、音频和视频。无需联网即可在手机上运行。
Meta
Llama 3.2 Vision
Meta Llama 3.2的多模态扩展版，增加了图像理解视觉编码器。11B版本可在单张8–12 GB VRAM GPU上运行。支持图像分析、OCR、图表阅读和视觉问答。
Mistral AI
Ministral
Mistral AI针对资源受限环境的轻量级模型系列。Ministral 8B的推理任务表现远超其规格，在6 GB VRAM GPU上即可运行。Ministral 3B专为手机和边缘设备设计。
Alibaba Cloud
Qwen 3.5
阿里巴巴2026年2月的旗舰新一代。Qwen 3.5立即登顶多项开源推理基准测试，GPQA Diamond得分88.4，SWE-bench Verified得分76.4。MoE架构提供8–19倍更高的解码吞吐量，支持200+语言。Apache 2.0许可证。
Moonshot AI
Kimi K2.5
Moonshot AI's 1-trillion parameter MoE flagship with 384 experts and a unique Agent Swarm architecture. Powers Cursor's Composer 2 IDE feature. Leads multiple coding and reasoning benchmarks including 93.33% on AIME 2026 and best Pass@1 on SWE-rebench. With 32B active parameters per token, requires datacenter infrastructure for self-hosting.
1T
最低~604.6 GB （Q4_K_M）
MiniMax
MiniMax M2.5
MiniMax's 230B MoE model scoring 80.2% on SWE-bench — technically outperforming Claude Opus 4.6 on this benchmark — at 22.7× lower cost on output tokens. Supports up to 1M token context and includes multimodal capabilities. Available under a modified MIT license requiring attribution. Primarily API-accessed but self-hostable on multi-GPU infrastructure.
最低~139.6 GB （Q4_K_M）
Xiaomi
MiMo-V2-Pro
Xiaomi's 1T+ MoE model originally launched anonymously as 'Hunter Alpha' in March 2026 before being revealed as MiMo-V2-Pro. Led by ex-DeepSeek engineer Luo Fuli. Was the largest model on OpenRouter at launch. Strong on PinchBench and ClawEval agentic benchmarks with ~42B active parameters per token. Available via API through Xiaomi and OpenRouter.
最低~604.6 GB （Q4_K_M）
DeepSeek
DeepSeek V3.2
DeepSeek 2026年初的V3进化版。V3.2（671B总参数，37B活跃参数）专注于改进长上下文推理、工具使用和智能体任务。在大多数基准测试中对标GPT-4o级别性能，MIT许可证完全开源。
671B/37B active685B
最低~414.4 GB （Q4_K_M）
NVIDIA
Nemotron Cascade 2
NVIDIA's hybrid Mamba-2 + MoE model, released March 2026 as a single checkpoint: Nemotron-Cascade-2-30B-A3B. 31.6B parameters resident, 3.5B active per token, and the second open-weight model to reach gold-medal level at IMO 2025, IOI and the ICPC World Finals. NVIDIA also open-sourced the post-training recipe. NVIDIA Open Model License, 1M-token context, first-party Ollama tag.
最低~19.9 GB （Q4_K_M）
OpenAI
GPT-OSS
OpenAI's open-weight family released under Apache 2.0 — their first substantial open-source release since GPT-2. Both sizes ship in MXFP4 precision, which is why they fit in far less memory than their parameter counts suggest: the 20B loads in ~14 GB and runs on a 16GB consumer card, while the 120B needs ~65 GB (dual RTX 3090 or a single 80GB A100). Signals OpenAI's strategic shift toward supporting on-premise enterprise deployments.
Mistral AI
Devstral
Mistral AI's coding specialists, spanning both generations. The original Devstral Small 2505 posts 46.8% on SWE-Bench Verified; its December 2025 successor Devstral Small 2 lifts that to 68.0% at the same 24B, which is the local-agent sweet spot — about 15 GB at Q4_K_M on one consumer card. The 123B flagship reaches 71.6%. Built for agentic software engineering: multi-file editing, repo navigation, and test-driven development. Apache 2.0 throughout.
Deep Cogito
Cogito v1
Deep Cogito's hybrid reasoning model family released March 2026. Unique 'think-or-not-think' architecture lets the model decide at inference time whether to use chain-of-thought reasoning based on query difficulty — saving compute on simple tasks while using deep reasoning for hard ones. Apache 2.0 licensed with sizes from 3B to 70B, all runnable on consumer hardware.
Moonshot AI
Kimi K2.5 / K2.6 / K2.7
Moonshot AI 2026年2月的突破之作。Kimi K2.5拥有1万亿总参数（32B活跃，384专家），创下HumanEval历史最高分99.0。在15万亿混合视觉和文本token上训练。具备独特的智能体群集架构。
最低~604.6 GB （Q4_K_M）
Moonshot AI
Kimi K3
Moonshot AI's July 2026 flagship and its largest model to date — a 2.8-trillion-parameter Mixture-of-Experts that landed in the July 16–27 window. K3 roughly doubles the expert count of Kimi K2.5 (1T) while keeping active parameters modest (~48B per token), pushing Moonshot's lead on agentic coding, tool use, and long-horizon reasoning further ahead of K2.6. Natively multimodal with a 256K context. At this scale it is a data-center / API model — self-hosting needs well over a terabyte of memory even at Q4, so most users reach it through the Kimi API or the smaller K2.x community builds.
最低~1691.3 GB （Q4_K_M）
Zhipu AI (Z.ai)
GLM-5 / GLM-5.1
Zhipu AI's (Z.ai) fifth-generation model series, released under MIT license for maximum flexibility. GLM-5 and GLM-5.1 excel at agentic workflows, tool calling, and long-context reasoning. The MIT license makes them one of the most commercially permissive frontier-class models available, with zero restrictions on commercial use or distribution. Both ship only as a 744B Mixture-of-Experts model (40B active, SWE-bench Verified leader) trained entirely on Huawei Ascend hardware: a data-center model, with no smaller consumer checkpoint. The 9B, 32B and 72B sizes listed here earlier were never released.
最低~6.2 GB （Q4_K_M）
Alibaba Cloud
Qwen 3.6
Alibaba's April 2026 follow-up to Qwen 3.5. As of June 15, 2026 only two tiers have been released — a 27B dense model (Apr 21–22) and a 35B-A3B MoE model (Apr 16) — both Apache 2.0, with native text/image/video input and ~256K context (extensible to ~1M via YaRN). Qwen 3.6-Plus/Plus-Preview/Max-Preview exist but are proprietary, API-only, and not listed here.
最低~17.6 GB （Q4_K_M）
DeepSeek
DeepSeek V4
DeepSeek's April 24, 2026 preview release, MIT licensed with a 1M-token context window. V4-Pro (1.6T total / 49B active) is server-class only — not runnable on consumer hardware even heavily quantized. V4-Flash (284B total / 13B active) is workstation-tier but local support is still WIP: llama.cpp has no merged upstream support as of June 2026 (only experimental forks), and Ollama's listings appear to route to cloud-hosted inference rather than a true local download. DeepSeek R2 remains unreleased/rumored as of June 15, 2026 and is intentionally not listed.
最低~172.3 GB （Q4_K_M）
Mistral AI
Mistral Small 4
Mistral AI's March 16, 2026 release unifying its former Magistral/Pixtral/Devstral lines into a single 119B-total / ~6.5B-active MoE model (128 experts, 4 active per token) with text + image input, reasoning, and agentic coding. Apache 2.0 licensed, 256K context. At Q4 it fits a single 24GB GPU (RTX 4090).
最低~72.6 GB （Q4_K_M）
IBM
IBM Granite 4.1
IBM's April 29, 2026 release. The 3B and 8B tiers are dense transformers; the 30B tier pivots back to a hybrid Mamba-Transformer (interleaved SSM + attention) design rather than 4.0's MoE. Apache 2.0 licensed, ISO 42001 certified, 128K context (extendable to 512K). All three sizes have official Ollama tags, spanning budget edge hardware to mid-range 24GB GPUs.
MiniMax
MiniMax M3
MiniMax M3 — a natively multimodal MoE at roughly 428B total and 23B active with a 1M-token context, under the minimax-community licence (June 2026). Multimodality is built into the single checkpoint: there is no separate vision release.
428B (23B active)230B (10B active)
最低~139.7 GB （Q4_K_M）
Mistral AI
Mistral Large 3
Mistral's open-weight flagship, released 2 December 2025: a 675B MoE with 41B active parameters, Apache-2.0, 256K context. Datacentre-scale to self-host; most readers will rent it.
最低~408.3 GB （Q4_K_M）
DeepSeek
DeepSeek V4.1
DeepSeek V4.1 — the Flash member shipped 10 September 2026 under MIT: a 552B-backbone MoE with a causal encoder-decoder design (8B active per input token, 16B per output token), a 196B-parameter Engram memory table and native image input. A V4.1 Pro was announced without a date and had not launched as of October 2026.
552B (16B active)1.6T (49B active)
最低~461.6 GB （Q4_K_M）
Alibaba Cloud
Qwen 3.7
Qwen 3.7 shipped as API-only Max and Plus tiers in May and June 2026. Alibaba skipped the generation for open weights and went straight to Qwen 3.8, so nothing in this family is locally runnable.
35B (3B active)
最低~21.9 GB （Q4_K_M）
Zhipu AI (Z.ai)
GLM-6
GLM-6 has not been released. Z.ai's current generation is GLM-5.3 (August 2026), alongside GLM-5.3-Flash. The listings here were withdrawn; they redirect to the nearest real model.
9B355B (32B active)
最低~6.2 GB （Q4_K_M）
Meta
Llama 4.5
Llama 4.5, reported as a June 2026 mid-cycle refresh of Llama 4 Scout and Maverick. UNVERIFIED: secondary sources disagree on whether this release happened, and no primary source has been read. Meta's confirmed 2026 open release is Muse Glimmer 30B (10 August, Apache-2.0).
最低~66.6 GB （Q4_K_M）
PrismML
Bonsai 27B
PREVIEW (July 2026). The first 27B-class model that runs on a phone. Bonsai 27B is PrismML's extreme low-bit rebuild of Qwen3.6-27B: the architecture is unchanged, but the weights are squeezed to roughly 1–1.7 bits each. The 1-bit build is just 3.9 GB — small enough for an iPhone — while the ternary build is 5.9 GB and aimed at laptops. Both are multimodal (text + images), carry a 262K-token context, and ship under Apache 2.0. PrismML reports the ternary build keeps more than 95% of the full-precision (FP16) baseline across a 15-benchmark suite and the 1-bit build more than 90%, running about 11 tok/s on an iPhone 17 Pro. Specs are from launch coverage — verify against the Hugging Face model cards.
Thinking Machines
Inkling
PREVIEW (July 2026). Thinking Machines Lab's first open-weights model — and the first ~1-trillion-parameter open model to natively accept text, image and audio input with a 1M-token context window. Inkling is a sparse multimodal Mixture-of-Experts transformer: 975B total parameters but only ~41B active per token (256 experts, top-6 routed plus 2 always-on shared experts). It was trained on 45 trillion tokens of text, images, audio and video, and reasons natively across modalities with seven selectable effort levels (none → max). Ships in full BF16 (~2 TB VRAM, Hopper+) and a well-calibrated NVFP4 4-bit build (~600 GB, Blackwell), both with MTP speculative-decoding layers for faster inference. Released under Apache 2.0 and pitched as a customization base to fine-tune via the company's Tinker platform — "built for customization, not leaderboard dominance." Datacenter-class hardware required; most users will reach it through a hosted API. Specs from launch coverage and the Hugging Face model card — verify before relying on them.
NVIDIA
Cosmos 3
PREVIEW (2026). NVIDIA's open frontier World Foundation Model (WFM) for Physical AI — the first fully open omnimodel that natively understands and generates text, images, video, ambient audio and robot/vehicle actions in a single system, with leading physics accuracy. Cosmos 3 is a two-tower Mixture-of-Transformers (MoT): a reasoner tower (an autoregressive vision-language "brain" that interprets images/video/text and reasons about motion, object interactions and physical cause-and-effect) plus a diffusion-based generator tower that predicts physics-aware future video and action sequences conditioned on that reasoning — so it "thinks before it acts." The two towers keep separate norm layers and MLPs but share multimodal attention. Trained on ~20 trillion tokens of multimodal data (nearly a billion images, 400M+ real and synthetic videos, plus audio and human/robot action data), it compresses physical-AI training and evaluation from months to days by generating synthetic worlds and predicting actions for robots and autonomous vehicles. Ships in three scales — Super (64B), Nano (16B) and Edge (4B) — under the permissive OpenMDW-1.1 license, on Hugging Face (nvidia/) and NVIDIA NGC. NVIDIA reports open-source state-of-the-art results on R-Bench, PAI-Bench, Physics-IQ and RoboLab; treat launch numbers as vendor-reported pending independent replication. This is a robotics / autonomous-vehicle / world-simulation model, not a chat LLM — verify specs on the model cards before relying on them.
Poolside
Poolside Laguna XS 2.1
Poolside's July 2026 open-weight coding model, and one of the few frontier-grade agentic coders small enough to run on a single consumer GPU. Laguna XS 2.1 is a 33B-total / 3B-active Mixture-of-Experts built for long-horizon software work — the 2.1 point release lifted SWE-Bench Verified from 69.9% to 70.9% and SWE-bench Multilingual from 57.7% to 63.1%. Architecturally it mixes sliding-window and global attention 3:1 across 40 layers, with sigmoid per-head gating and an FP8 KV cache, which is how it holds a 256K context without a datacenter behind it.
最低~20.7 GB （Q4_K_M）
NVIDIA
Nemotron 3 Super
NVIDIA's 120B open Mixture-of-Experts, activating just 12B parameters per token for compute efficiency on multi-agent workloads. Like the rest of the Nemotron line it generates a reasoning trace before its final answer, and NVIDIA publishes open weights alongside training data and recipes rather than weights alone. Positioned for collaborative agents and high-volume automation such as IT ticket triage. First-party Ollama tags exist at Q4_K_M, Q8_0 and BF16, so this is genuinely self-hostable — on a multi-GPU box rather than a laptop.
最低~73.3 GB （Q4_K_M）
NVIDIA
Nemotron 3 Ultra
NVIDIA Nemotron 3 系列的顶配：一个 5500 亿参数的混合模型，交错使用 Mamba-2 层、潜在专家混合（LatentMoE）层和少量注意力层，每个 token 激活 550 亿参数。采用 NVFP4 方案预训练，以 BF16 和 NVFP4 两种格式在 OpenMDW-1.1 许可下发布。NVIDIA 给出的 BF16 检查点最低配置是一台 8 卡 H200 或 B200 服务器。
最低~339.2 GB （Q4_K_M）
Zhipu AI (Z.ai)
GLM-5.2
Z.ai's June 2026 flagship and, by Artificial Analysis' ranking at the time, the strongest open-weight model available — fourth overall including closed models. A 744B-parameter Mixture-of-Experts with roughly 40B active per token and a genuinely usable 1M-token input context (output up to 131,072). Two selectable reasoning-effort levels. Released in stages between 13 and 17 June 2026: the Coding Plan first, then open weights on Hugging Face, then documentation. The MIT license is the notable part — full commercial use, modification and redistribution on a frontier-scale model.
最低~450 GB （Q4_K_M）
MiniMax
MiniMax M2.7
MiniMax's March 2026 agentic MoE: 230B total parameters with 10B active per token across 256 experts — a 4.3% activation rate — over 62 layers, using multi-head causal self-attention with RoPE. Weights are published openly on Hugging Face and GitHub, so this is self-hostable rather than API-only, though the scale still implies a multi-GPU box. NVIDIA has upstreamed fused QK RMS Norm and FP8 MoE kernels from TensorRT-LLM for it. ⚠️ Released under a NON-COMMERCIAL license — commercial deployment requires a separate agreement with MiniMax.
最低~139.7 GB （Q4_K_M）
IBM
IBM Granite 4.0
The Granite 4.0 generation, listed here for the one tier that has no Granite 4.1 counterpart: the 32B-total / 9B-active hybrid mamba-2 build. Granite models are finetuned from their base models on permissively-licensed open instruction datasets plus internally-collected synthetic data, with improved instruction-following and tool-calling aimed squarely at enterprise RAG and agent workloads. Apache 2.0. For the smaller sizes, see Granite 4.1 — those tiers are newer and supersede their 4.0 equivalents.
最低~20.1 GB （Q4_K_M）
Hugging Face
SmolLM3
Hugging Face's fully-open 3B: weights, training data, recipe and 100+ intermediate checkpoints all published. Dual-mode — a deep-thinking path for hard problems and a fast path for everything else — with 128K context and six natively supported languages. The reference model for anyone fine-tuning at the small end, which makes it the natural companion to the datasets hub.
最低~2.7 GB （Q4_K_M）
WeiboAI
VibeThinker
MIT-licensed reasoning fine-tunes that argue the case for small models: VibeThinker-3B scores 94.3 on AIME26, matching what DeepSeek V3.2 reaches at 671B. Trained with the Spectrum-to-Signal pipeline — explore solution diversity during SFT, then reinforce the correct signal with RL. Vendor-reported and contested; treat the headline comparison as a claim about a narrow benchmark family, not general capability.
Mistral AI
Magistral Small
Mistral's reasoning line: Mistral Small with long chain-of-thought added via SFT on Magistral Medium traces plus RL. Apache 2.0, vision-capable, and sized so the whole thing fits a single RTX 4090 or a 32 GB MacBook once quantized — which is the point. Reasoning traces degrade past roughly 40K tokens even though the context window runs to 128K.
最低~15.3 GB （Q4_K_M）
Mistral AI
Mistral Small 3.2
A reliability release rather than a capability one. Same 24B architecture as Small 3.1, with instruction-following, function-calling robustness and output stability fixed — Arena Hard more than doubles (19.56% to 43.10%) and the infinite-repetition rate halves. If you were running 3.1, this is a drop-in upgrade with no new hardware cost.
最低~15 GB （Q4_K_M）
NVIDIA
Nemotron 3 Nano Omni
The first mainstream omni-modal open model that a single consumer GPU can actually hold: text, image, speech and video in, text out, at 3B active parameters from a 30B MoE. 260K context under the NVIDIA Open Model License, which permits commercial use. The efficiency ratio here is the story — video understanding at roughly the per-token cost of a 3B dense model.
最低~18.9 GB （Q4_K_M）
Alibaba Cloud
Qwen3.8
Alibaba's August 2026 wave. The 27B is a natively multimodal dense model under Apache 2.0 with a 262K context that extends to 1M via YaRN — small enough for one consumer card, which is the combination most of this site's visitors are looking for. Qwen3.8-Max sits at the other extreme: 2.4T parameters with roughly 95B active, open-weighted under its own licence. Qwen3.8-Flash-Next is neither — an experimental preview of the Qwen4 architecture, 180B resident but only 6B active, under the Qwen Community License rather than Apache.
最低~17.6 GB （Q4_K_M）
Xiaomi
MiMo-V2.5
Xiaomi's MIT-licensed MoE pair, and the clearest licence story at the trillion-parameter tier — MIT, not a bespoke open-weight agreement with revenue thresholds. Both carry 1M-token context and agentic multimodal capability. Consistently omitted from best-of roundups despite matching Kimi K2.6 on aggregate index scores.
最低~188 GB （Q4_K_M）
Xiaomi
MiMo-V2.6
小米 2026 年 9 月发布的旗舰，是 MiMo-V2.5 与 MiMo-V2-Pro 的后继系列。唯一发布的检查点 MiMo-V2.6-Pro 是一个 1.02 万亿参数的稀疏 MoE，每个 token 激活 420 亿参数，支持文本、图像、视频和音频输入，上下文 100 万 token，采用 MIT 许可。所有专家都必须常驻内存，因此这是数据中心级模型——小米官方通过 SGLang 和 vLLM 在多 GPU 服务器上部署。
最低~6.2 GB （Q4_K_M）
Mistral AI
Ministral 3
The edge tier of the Mistral 3 family — 3B, 8B and 14B dense, every size shipped in base, instruct and reasoning variants with image understanding, all Apache 2.0. Built for laptops, phones, drones and anywhere a datacenter is not. The 3B runs in about 4 GB of video memory; the 14B reasoning variant scores 85% on AIME 2025.
Nex-AGI
Nex-N2.5
The September 2026 agentic family from Nex-AGI, an innovation alliance initiated by the Shanghai Innovation Institute. Three sizes, all post-trains of other labs' base weights: mini and Pro continue the multimodal Qwen3.5 lineage of Nex-N2, while Max is the alliance's first trillion-scale post-train, built on a 1.6T text-only DeepSeek-V4-Pro base. Vision is treated as an action interface rather than an input mode — the models drive computers and browsers and check their own work from what appears on screen. Max takes BrowseComp at 92.6, ahead of every model in Nex's own comparison table. Only the 35B mini is realistically self-hostable.
最低~21.9 GB （Q4_K_M）
Nex-AGI
Nex-N2
Nex-AGI's Apache-2.0 agentic pair, post-trained on Qwen3.5 and built around what Nex calls Agentic Thinking: Adaptive Thinking lets the model choose when and how deeply to reason, Coherent Thinking keeps one reasoning style across chat, tool calls and terminal work. Pro reaches 75.3 on Terminal-Bench 2.1 and 90.7 on GPQA Diamond. The 35B mini is the size that matters for local use — 21.9 GB at Q4_K_M with 3B active parameters per token. This is the generation whose parameter counts Nex states outright.
最低~21.9 GB （Q4_K_M）
Nex-AGI
Nex-N1
Nex-AGI's first agentic series, and the one that set the pattern: take someone else's open base weights, post-train them for tool use and autonomy, publish the result. The flagship is a DeepSeek-V3.1 post-train scoring 80.2 on tau2-Bench and 70.6 on SWE-bench Verified. Superseded by Nex-N2, but kept because the question 'what does a 671B agentic post-train need?' has a factual answer that does not change when the model stops being recommended.
最低~405.9 GB （Q4_K_M）
Zhipu AI (Z.ai)
GLM-5.3-Flash
The first natively multimodal model in the GLM-5 series, and the one Z.ai actually released under MIT — 320B total parameters with 18B active, text plus image and video in, a 1M-token context, and hybrid attention (linear attention for local dependencies, sparse attention for distant retrieval). Z.ai claims 63.4 on DeepSWE against GLM-5.2's 46.2. Note this is NOT the GLM-5.3 flagship: that is a separate 753B model under a bespoke licence with a hyperscaler security-review clause.
最低~194 GB （Q4_K_M）
Zhipu AI (Z.ai)
GLM-5.3
Z.ai 于 2026 年 8 月发布的 7530 亿参数旗舰。它与 GLM-5.2 共用基座模型，所有提升都来自后训练；模型卡重点展示智能体编程与网络安全能力。仅支持文本，上下文 100 万 token，采用自有许可：类 MIT，但年收入超过 100 亿美元的“模型即服务”运营方在商用前须通过 Z.ai 的安全审查。原生多模态的 320B 版本请见 GLM-5.3-Flash。
最低~455.6 GB （Q4_K_M）
Aleph Alpha
Kolibri
Kolibri 1 是 Aleph Alpha 面向德语和英语的混合专家（MoE）推理模型，于 2026 年 10 月发布，权重采用 Apache-2.0 许可证。它从零开始训练，支持可调节的推理强度和工具调用，并以大部分滑动窗口注意力的设计降低长上下文的缓存成本。内存方面的代价是实实在在的：所有专家都必须驻留内存，因此它是服务器和大内存工作站模型，而不是笔记本模型。
最低~48.3 GB （Q4_K_M）
Arcee AI
Trinity
Trinity 是 Arcee AI 的开放权重稀疏混合专家模型系列。Trinity Mini（总参数 26B，约 3B 激活）是适合本地运行的成员，有官方 GGUF；Trinity-Large-Thinking（总参数约 398B，约 13B 激活）是面向多 GPU 服务器的推理与智能体模型。整个系列已于 2026 年 5 月 29 日从 Apache-2.0 改为 OpenMDW-1.1，因此较早出现的 Apache 说法已经过时。
最低~16.7 GB （Q4_K_M）
NVIDIA
Nemotron 3 Nano
Nemotron 3 Nano 4B 是 NVIDIA 的紧凑文本模型，由 Nemotron-Nano-9B-v2 压缩为 Mamba-2 与注意力的混合架构，仅有四个注意力层。它面向边缘设备和 RTX 电脑，提供可切换的推理模式，支持最高 262K 上下文，并提供官方 Q4_K_M GGUF 和 Ollama 标签。
4B
最低~3.6 GB （Q4_K_M）
ByteDance Seed
Seed-OSS
Seed-OSS-36B 是字节跳动 Seed 的稠密 36B 开放权重推理模型，采用 Apache-2.0 许可证，具备可控的思考预算、工具调用解析器支持和原生 512K 上下文窗口。它发布于 2025 年 8 月，作为受支持的旧版选项保留在库中；长上下文是模型的能力，并不意味着单张消费级 GPU 就能支撑。
最低~22.6 GB （Q4_K_M）
Allen Institute for AI (Ai2)
OLMoE
OLMoE-1B-7B 是 Ai2 的小型、文档完备的混合专家研究模型，每层 64 个专家，总参数 6.9B，每个 token 约激活 1.3B。2025 年 1 月的 Instruct 版本提供官方 GGUF 文件，适合手机、平板和 Apple Silicon 等轻量场景，但它以英语为主、上下文较短且较旧，因此列为小众的旧版条目。
7B (1.3B active)
最低~5 GB （Q4_K_M）
Z.ai
GLM-4.5
GLM-4.5-Air 是 Z.ai 的混合推理混合专家模型：总参数 106B，激活 12B（支持思考和非思考模式），权重采用 MIT 许可证。它发布于 2025 年中，之后被更新的 GLM 系列取代，但仍是大内存本地部署的热门选择；它被列为旧版，并链接到更新的 GLM 条目。
106B (12B active)
最低~73.8 GB （Q4_K_M）
Cohere
North Mini Code
Cohere's first model aimed at developers rather than enterprises, and the opening entry in their North family of code agents. A 30B sparse MoE activating ~3B per token, Apache 2.0, 256K context and 64K maximum generation — small enough to run locally while posting a Coding Index of 33.4, ahead of several open models many times its size.
最低~18.9 GB （Q4_K_M）
MBZUAI Institute of Foundation Models
K2 Horizon
来自 MBZUAI 基础模型研究院（Institute of Foundation Models）的完全开放模型家族——权重、训练数据、训练方案和代码均以 Apache 2.0 许可发布，并公开了中间检查点。共三个尺寸：面向小型设备的 3.7B 和 7B 稠密模型，以及每个 token 约激活 4B 参数的 36B 混合专家模型。三者均原生支持 512K 上下文。在独立的 Artificial Analysis 智能指数中分别得到 16、21 和 25 分，核查时为各自尺寸级别中的最高分。
最低~3.9 GB （Q4_K_M）
IBM
IBM Granite 4.2
IBM's first family of dense decoder-only reasoning models, released August 2026 in 3B, 8B and 30B. Granite 4.2 drops the hybrid Mamba-2 MoE design of Granite 4.0 and continues the dense line Granite 4.1 started. Each size is pretrained from scratch on roughly 15T tokens with a five-phase context extension to 512K, then RL-post-trained for tool calling, RAG, structured JSON and an explicit thinking mode. All three are Apache 2.0.
OpenBMB
MiniCPM-V
OpenBMB's pocket-sized multimodal line, backed by ModelBest and Tsinghua University, built around one goal: image and video understanding that actually runs on a phone. Version 4.5 is the 8B that beat much larger proprietary models on OpenCompass; version 4.6 is the 1.3B successor that cut visual-encoding FLOPs by more than half and ships on iOS, Android and HarmonyOS.
DeepSeek
DeepSeek-OCR
A 3B vision-language model built for one job — turning document pages into structured text — and a demonstration of "contexts optical compression": rendering text as an image and decoding it costs roughly 10x fewer tokens at near-lossless fidelity, or 20x at about 60% precision. A DeepEncoder compresses each page into vision tokens which a DeepSeek-3B MoE decoder reads, activating only ~570M parameters. A second generation, DeepSeek-OCR-2, followed in January 2026.
最低~2.6 GB （Q4_K_M）
Liquid AI
LFM2.5
Liquid AI's on-device line, built around their hybrid LIV convolution architecture rather than a stack of attention blocks. The 8B-A1B is the flagship: a sparse MoE small enough for phones, laptops and robots, reasoning-only (it emits an explicit chain of thought before answering), and trained on 38T tokens with large-scale RL. Liquid claims tool-calling quality comparable to models up to 4x its active size.
最低~2.4 GB （Q4_K_M）
Meta (Superintelligence Labs)
Muse Glimmer
Meta重返开放权重，也是自Llama 4以来的第一次。一个29.6B稠密模型，从闭源的Muse Spark蒸馏而来，配备自有的感知编码器，专为在单张消费级显卡上完成智能体任务而构建：多步推理、工具调用和错误恢复集中在一个无需联网即可运行的检查点中。
最低~18.7 GB （Q4_K_M）
NVIDIA
Nemotron 3.5
NVIDIA的智能体产品线，构建在一套混合架构上，将Mamba-2状态空间层与MoE模块交错排列，只夹杂少数注意力层。只有注意力层保留KV缓存，这正是它能在这一参数规模下宣称1M token上下文的原因。
最低~18.9 GB （Q4_K_M）
Zhipu AI (Z.ai)
GLM-4.6V
Z.ai的视觉语言产品线。Flash成员是9B的高效变体，以极小的算力代价共享GLM-4.6V的多模态能力，小到可以装进8 GB显卡。
最低~6.2 GB （Q4_K_M）
Ai2 (Allen Institute for AI)
Olmo 3
完全开放，而不只是开放权重：Ai2以Apache-2.0许可公开权重、训练数据（Dolma 3）、代码和中间检查点。每个尺寸都提供Base、Instruct和Think三种变体，因此当训练过程本身必须可审计时，应当首选这个系列。
SpeakLeash / ACK Cyfronet AGH
Bielik
波兰的开放LLM，在波兰大规模计算基础设施（PLGrid）上训练，语料由SpeakLeash团队整理。v3版把分词器换成了专为波兰语构建的APT4，这也是它在波兰语文本上能比同等规模的通用多语言模型发挥出更多性能的原因。
最低~7.4 GB （Q4_K_M）
Unbabel / Instituto Superior Técnico / INESC-ID
EuroLLM
在巴塞罗那超级计算中心的MareNostrum 5超级计算机上从零训练，覆盖欧盟全部24种官方语言。这是欧洲对多语言支持只是附带功能的那类模型给出的回应：欧盟语言是训练目标，而不是副产品。
EPFL / ETH Zurich / CSCS
Apertus
瑞士的开放模型，来自EPFL、苏黎世联邦理工学院和瑞士国家超级计算中心。在约15万亿token上训练，覆盖1000多种语言，权重、数据和流程均以Apache-2.0许可公开——把欧洲的主权论点落到了实处。
PrismML
Ternary Bonsai 2
PrismML 的第二代三值模型系列：四个视觉语言模型，权重只取 −1、0 或 +1，以打包格式存储，体积约为 FP16 的九分之一。2B、4B 和 9B 由 Qwen3.5 重建，27B 由 Qwen3.8-27B 重建，因此都保留 262,144 token 上下文和混合注意力结构（每三层线性注意力配一层全注意力），KV 缓存也因此很小。它们只以这些打包格式发布——没有 Q4_K_M 或 Q8_0 版本——最小的文件需要 PrismML 自己的 llama.cpp 分支或 mlx-vlm 才能加载。Apache 2.0。27B 取代了 7 月基于 Qwen3.6 的 Bonsai 27B。
Meta
Muse Spark Open
Meta 于 2026 年 10 月 15 日发布的 Muse Spark 开放权重版本：总参数 816B、激活 42B 的混合专家模型，输入支持文本、图像、视频和音频，上下文 1M token。所有专家都必须驻留内存，因此即使是 Meta 的 4 位（NVFP4）检查点也要下载 533 GB，Meta 要求至少 700 GB GPU 显存才能运行。它是多 GPU 服务器模型，而不是桌面模型。如果你想在一张消费级显卡上使用 Meta 的权重，这条产品线的本地成员是 Muse Glimmer 30B；本库中其他单 GPU 替代选择有 Ternary Bonsai 2 27B 和 Qwen3.8 27B。
最低~497.1 GB （Q4_K_M）
JetBrains
Mellum2.1
JetBrains 的开放编程模型。Mellum2.1 Thinking 是 12B 参数、2.5B 激活的混合专家模型，主要通过在真实代码仓库中的强化学习进行后训练（使用 shell 和文件编辑工具，测试通过即获得奖励），用于智能体式软件工程和作为子智能体。Apache 2.0，上下文 131,072 token。Q4_K_M 约 8 GB，是少数能装进中端笔记本的智能体编程模型之一。
最低~8.9 GB （Q4_K_M）
Mistral AI
Mistral Medium 3.5
Mistral 首个合并式旗舰：一个 128B 稠密模型，把推理线（Magistral）和编程线（Devstral 2）合并进同一个检查点，支持图像输入和 256K 上下文。稠密意味着每个 token 都要读取全部参数，所以每 token 比同等规模的混合专家模型慢，但 Q4_K_M 约 78 GB，可放进 128 GB 的 Mac 或两张 48 GB 显卡。它的“Modified MIT”许可证对月收入超过 2000 万美元的公司不授予任何权利。
最低~77.9 GB （Q4_K_M）
Bespoke Labs
Nimble
Bespoke Labs 的 9B 决策模型，基于 Qwen3.5-9B 的 LoRA，使用对比数据对训练。是 Ollama 库中的默认决策模型。
最低~6.6 GB （Q4_K_M）
Together AI
Tev1 (experimental)
Together AI 受 Jev 启发的实验性决策模型，是对 Qwen3.5 的监督微调，从结构化决策中选出一个选项。可在 Ollama 中使用。
EldanRing
Winnow
基于 Gemma 4 微调的类型化决策模型，以 GGUF 发布，配套基于 llama.cpp 的服务器，同一个已加载的模型同时提供 /v1/systemone 与聊天、视觉能力。
最低~5.6 GB （Q4_K_M）
Mapika
Decider
基于 Qwen3.5 基座（0.8B 至 35B-A3B）的开源决策模型系列，接口统一，每个问题最多支持 255 个选项，并配有兼容 TypeSafe 的服务器。
最低~1.3 GB （Q4_K_M）
alibiserikbay
JevK5
基于 Qwen3.5-4B 的 4B 决策模型，带有 TypeSafe 风格的服务器，并提供 GGUF 版本，可在 NVIDIA、AMD、Intel、Apple GPU 或 CPU 上通过 llama.cpp 运行。
最低~3.6 GB （Q4_K_M）
InternLM
Intern-Decision
InternLM 基于 Qwen3.5 的多模态结构化决策模型，一次前向传播即可回答模式中的每个问题。
最低~3.6 GB （Q4_K_M）
denis-pplx
AutoJev
基于 Qwen3.8-27B 的 27B 多模态决策模型，截至 2026 年 9 月 28 日是社区 Decision Index 上得分最高的可下载模型。
最低~17.6 GB （Q4_K_M）
ConvAI Innovations
Laya
基于编码器（ModernBERT / mmBERT）的小型决策模型，用于 100 多种语言的快速路由与安全护栏，通过 pip 安装。
最低~1.1 GB （Q4_K_M）
Cloudflare
Clef
Cloudflare 的决策模型 Clef 27B 与 Clef-Flash 9B：输入带类型的问题，一次评分即输出每个选项的概率，不生成文本。两者都在 Ollama 库中，并支持图像输入。
Surogate
Rune
Surogate 基于 Gemma 4 26B-A4B 的决策模型（混合专家，文本与图像）。在 2026 年 9 月 28 日的社区 Decision Index 0.2.1 快照中是得分最高的可下载模型。
最低~16.4 GB （Q4_K_M）
Manjunath Janardhan
OpenDecider
从经过校准的教师模型蒸馏而来的开源决策模型：400M 编码器（nano）和基于 Qwen3-4B 的 4B LoRA（small），通过 pip 安装，由兼容 /v1/systemone 的 opendecider serve 提供服务。更大的 30B 与 80B MoE 版本需要 NVIDIA GPU，未在此列出。
Hanno Labs
Bosun
Hanno Labs 的小型决策模型：基于 Qwen3-0.6B 与 Qwen3-1.7B 的 rank-16 LoRA，带 256 个学习得到的决策 token，单个问题最多 255 个候选项。由兼容 Jev 的服务器通过 /v1/systemone 提供服务。
Liquid AI
d1
Liquid AI 面向边缘设备的决策模型：d1-3B（文本与图像，基于 LFM2.5-VL-3B）和实验性的 d1-omni-600M（文本加图像或音频）。通过 Transformers 加载的 Python 库，而非服务器；采用 LFM Open License v1.0 许可。
HAL-X AI
THX-01
HAL-X AI 的 322M 多语言决策模型：mmBERT-base 编码器加决策头，在 18 种语言上进行了后训练。除 choice、是/否和 score 外，还能从文档中返回数字、原文摘录和引用。
最低~1 GB （Q4_K_M）
Alibaba Cloud已被取代
Qwen 3.5 (Withdrawn Listings)
Alibaba's February 2026 flagship generation. Qwen 3.5 immediately topped multiple open-source reasoning benchmarks, scoring 88.4 on GPQA Diamond and 76.4 on SWE-bench Verified. The MoE architecture delivers 8–19× higher decoding throughput vs earlier Qwen3 while supporting 200+ languages. Apache 2.0 licensed for full commercial use.
7B14B32B122B/10B active
最低~5 GB （Q4_K_M）
Google DeepMind已被取代
Gemma 4 (Withdrawn Listings)
Google DeepMind's breakthrough open-weight model. Gemma 4 27B delivers GPT-4 level performance at 14 GB VRAM, making frontier-quality AI accessible on consumer GPUs. Features a hybrid architecture with interleaved local and global attention, multi-image understanding, and 128k context window. Achieves 85 tokens/second on RTX 4090.
4B12B27B
最低~3.2 GB （Q4_K_M）
DeepSeek已被取代
DeepSeek V4 Pro (Withdrawn Listings)
Rank #1 open-weight model on every major benchmark. DeepSeek V4 Pro uses a massive 1.6 trillion parameter Mixture-of-Experts architecture with only 49 billion parameters active per token — giving frontier-tier intelligence at a fraction of the compute cost. Dominates coding, math, reasoning, and multilingual tasks.
1.6T / 49B activeAPI
最低~966.8 GB （Q4_K_M）
Alibaba Cloud已被取代
Qwen 3.5 (Withdrawn Listings)
Alibaba's most ambitious open-weight model release. Qwen 3.5 (397B) is a natively multimodal model supporting text, vision, and video understanding in a single architecture. Ranks among the top 3 open-weight models globally and competes directly with GPT-4o and Gemini 2.0 Pro on multimodal benchmarks.
14B72B397B
最低~9.3 GB （Q4_K_M）
DeepSeek已被取代
DeepSeek V3.2-Exp
DeepSeek-V3.2-Exp, the experimental sparse-attention snapshot released in September 2025 ahead of the V3.2 line.
685B (37B active)
最低~414.4 GB （Q4_K_M）
模型库中也收录了决策模型（Jev 风格的 System One：它们用每个选项的概率而不是文本来回答带类型的问题）。可以先看决策模型指南，或为其生成请求；这篇说明文章解释了基本概念。它们被有意排除在排名之外。