开放权重 · 怎么选Open weights · how to choose
开放权重，该怎么选How to choose open-weight models先选家族与规格，再选量化版——两步都有出处Pick the family and size first, the quantisation second — both steps carry sources
你想在自己电脑上跑一个大模型。问题分两层：第一层，开放权重的家族、代次、尺寸档、许可各不相同，该选哪个；第二层，同一个模型有几十个量化版，该下哪一个。本站把两层分开答：家族谱系卡讲第一层（出品方 / 代次 / 尺寸 / 许可 / 官方声明），系列页讲第二层（你的显存能跑哪档、该下多大体积）。不评模型强弱，也不排名。You want to run a large model on your own machine. The question has two layers: first, which family, generation, size tier and licence; second, which of dozens of quantisations to download. This site answers them separately: family lineage cards cover layer one (vendor / generations / sizes / licences / official claims), and series pages cover layer two (which tier fits your VRAM, how big the download is). It does not rank models against each other.
收录范围：开放权重（含蒸馏）——可本地部署的模型都收，无论它是原生训练还是从别家模型蒸馏而来（HF 上有公开权重的 MoE 与 dense 模型）。闭源 / API-only 模型不在量化收录范围内——它们没有量化产物，下载和选型的问题不存在。本站另设闭源谱系卡做对照：只记官方规格快照与复核周期，不采价格、不排名。Scope: open weights (distillation included) — any model you can run locally, whether trained from scratch or distilled from another model (MoE and dense models with public weights on HF). Closed-source / API-only models are out of the quantisation scope — they have no quantisations, so the download and selection questions do not arise. The site keeps a separate set of closed-model lineage cards for contrast: official spec snapshots with a review cycle, no prices, no rankings.
先说清楚：这个站的数据怎么来的First, where this site's data comes from
直接从 Hugging Face 公开 API 读取的客观数据：文件体积（字节）、上下文长度、架构、许可、下载量、更新时间。Objective data read straight from the Hugging Face public API: file sizes in bytes, context length, architecture, licence, downloads, last updated.
由采集字段按公开规则算出：每权重比特数（依 llama.cpp 命名标准）、档位归属、显存需求区间。推导规则在Computed from the collected fields by published rules: bits per weight (per the llama.cpp naming convention), tier assignment, VRAM requirement range. The rules are fully published in METHODOLOGY.md 完整公开，任何人可复算。 — anyone can recompute them.
哪个量化质量更高、哪个更快、哪个模型更强——这些需要实测，本站不跑 benchmark，所以不装作知道。凡我们给出的主观判断，一律显式标注「我们的口径」。Which quantisation is higher quality, which is faster, which model is stronger — these need measurement. This site does not run benchmarks, so it does not pretend to know. Every subjective judgement we do offer is explicitly marked as our own reading.
本站This site does 不下载、不托管任何模型权重文件not download or host any model weight files。所有数字都来自公开元数据，点击即可回到发布页核对。. Every figure comes from public metadata, and clicking it takes you back to the release page to verify.
采集时间Collected: 2026-10-30（数据快照，非实时） (a snapshot, not live)
家族谱系Model families 第一层：出品方 · 代次 · 尺寸档 · 许可 · 官方声明layer one: vendor · generations · size tiers · licences · official claims
每个开放权重家族一张卡：它是什么来头、现在哪一代、有哪些尺寸档、许可能不能商用、官方声明过什么能力（声明全部挂官方出处）。家族内哪一代有本站量化对比，卡上直接给出入口。One card per open-weight family: where it comes from, which generation is current, what size tiers exist, whether the licence allows commercial use, and what the vendor officially claims (every claim links to its official source). Where a generation has quantisation coverage on this site, the card links straight to it.
Command ACommand A
CohereCohere当前代current Command A+Command A+ · 2026-05
Cohere 的企业向 agentic 家族：A+ 放开了 Apache-2.0（上一代还是门控），主打企业场景与多语言。尺寸在 218B 级，本地部署门槛高。本站未收其量化对比（GGUF 生态太薄，见下）。Cohere's enterprise agentic line: A+ opened up under Apache-2.0 (the previous generation was gated), pitched at enterprise and multilingual work. At 218B scale the local-deployment bar is high. This site carries no quantisation comparison (GGUF ecosystem too thin — see below).
未收量化对比，卡内说明原因 →No quantisation coverage — see the card →
DeepSeekDeepSeek
深度求索 DeepSeekDeepSeek当前代current V4.1-FlashV4.1-Flash · 2026-09
MIT 许可 + KV cache 压缩路线鲜明；量化生态活跃（Redis 作者、llama.cpp 生态参与者 antirez 维护过其量化版，本站系列页有记录）。适合追新、能接受大模型的部署者。MIT-licensed with a distinctive KV-cache-compression roadmap; an active quantisation ecosystem (antirez, author of Redis and a llama.cpp ecosystem participant, maintains a quantised build; recorded on the series page). For deployers who track the frontier and can host large models.
看量化对比（DeepSeek V4）→Quantisation coverage (DeepSeek V4) →
GemmaGemma
Google DeepMindGoogle DeepMind当前代current Gemma 4Gemma 4 · 2026-03
尺寸档最全的家族（E2B 到 31B），端侧到服务器一条线；许可是 Apache-2.0 + 使用政策组合，下载前读政策原文。本站收录的 gemma-4 系列覆盖其中三档。The family with the widest size spread (E2B to 31B), one line from edge to server. Licensing is Apache-2.0 plus a use policy — read the policy text before download. This site's gemma-4 series covers three of the tiers.
看量化对比（Gemma 4）→Quantisation coverage (Gemma 4) →
GLM 智谱GLM (Z.ai)
智谱 AI · Z.aiZ.ai (Zhipu AI)当前代current GLM-5.3GLM-5.3 · 2026-08
中文 agent 与代码强项、迭代节奏极快（半年四代）。大杯与 Flash 杯许可不同——本地部署优先 MIT 的 Flash 杯。本站收录的 glm-5 系列在 GLM-5.3 上。Strong at Chinese agentic work and code, with a very fast iteration cadence. Flagship and Flash tiers carry different licences — for local deployment, prefer the MIT Flash tier. This site's glm-5 series tracks GLM-5.3.
看量化对比（GLM 5）→Quantisation coverage (GLM 5) →
gpt-ossgpt-oss
OpenAIOpenAI当前代current gpt-ossgpt-oss · 2025-08
OpenAI 首次放出的开放权重 MoE：Apache-2.0、官方就按「量化后交付」设计（MXFP4），单卡可跑是它的主打场景。本站收录 gpt-oss 系列的量化对比。OpenAI's first open-weight MoE line: Apache-2.0, designed to ship quantized (MXFP4) from the start, single-GPU operation is its headline scenario. This site covers the gpt-oss quantisation comparison.
看量化对比（gpt-oss）→Quantisation coverage (gpt-oss) →
KAT CoderKAT Coder
Kwaipilot（快手）Kwaipilot (Kuaishou)当前代current KAT-Coder-V2.5KAT-Coder-V2.5 · 2026-07 派生自 Qwen3.6-35B-A3Bderived from Qwen3.6-35B-A3B
快手 Kwaipilot 的代码家族：API 版闭源、Dev 版开放权重。它是本站收录的「派生模型」代表——官方自己声明基座是 Qwen3.6-35B-A3B，选它等于选 Qwen 谱系 + 快手代码后训练。Kuaishou Kwaipilot's coding line: the API build stays closed, the Dev build is open-weight. It is this site's representative "derived model" — the vendor itself declares Qwen3.6-35B-A3B as its base, so choosing it means choosing the Qwen lineage plus Kuaishou's code post-training.
看量化对比（KAT Coder V2.5）→Quantisation coverage (KAT Coder V2.5) →
KimiKimi
月之暗面（Moonshot AI）Moonshot AI当前代current Kimi K3Kimi K3 · 2026-06
国产开放权重旗舰：K3 是官方口径的首个开放 3T 级模型，许可是自定义 Kimi K3 License。对话产品「Kimi 智能助手」在网信办首批备案（Beijing-MoonShot-20231016）；模型权重本身与备案是两回事，卡上分开写。The domestic open-weight flagship: K3 is, per the vendor, the first open 3T-class model, under a custom Kimi K3 License. The consumer chat product was filed in the CAC's first batch (Beijing-MoonShot-20231016); model weights and service filing are different things — the card keeps them apart.
看量化对比（Kimi K3）→Quantisation coverage (Kimi K3) →
LagunaLaguna
PoolsidePoolside当前代current Laguna 2.1 / XS.2Laguna 2.1 / XS.2 · 2026-08
代码专用家族里许可最宽的一族（OpenMDW-1.1 完全宽松）；主打「本地机器跑 agentic coding」，激活 3B 是它的本地性来源。本站收录 Laguna XS-2.1 的量化对比。The most permissively licensed coding line (OpenMDW-1.1, fully permissive); its pitch is agentic coding on a local machine, with 3B active parameters as the source of that locality. This site covers Laguna XS-2.1 quantisations.
看量化对比（Laguna 2.1）→Quantisation coverage (Laguna 2.1) →
LlamaLlama
MetaMeta当前代current Llama 4Llama 4 · 2025-04
全球生态最大的开放权重家族：微调、量化、工具链覆盖最全，但许可是带归属条款（Built with Llama）与规模条款（7 亿 MAU）的社区许可，商用前读全文。The largest open-weight ecosystem by far — fine-tunes, quantisations and tooling are everywhere — but the licence is a community licence with attribution ("Built with Llama") and scale (700M MAU) clauses. Read it fully before commercial use.
未收量化对比，卡内说明原因 →No quantisation coverage — see the card →
MiniCPMMiniCPM
面壁智能 · OpenBMBOpenBMB (ModelBest)当前代current MiniCPM5MiniCPM5 · 2026-08
国产端侧小杯家族：尺寸小、官方连训练数据一起开源，按「自己练一遍」路线能拿到完整起点的中文家族（训练数据也一起放出）。面壁 ChatGLM 之外的另一条国产备案线（MiniCPM 备案编号 Beijing-MiniCPM-202503180058）。The domestic on-device small-model family: small sizes plus fully open training data — a Chinese family that gives you for the "train one yourself" route. Alongside Zhipu, another filed domestic line (MiniCPM filing number Beijing-MiniCPM-202503180058).
看量化对比（MiniCPM 5）→Quantisation coverage (MiniCPM 5) →
MistralMistral
Mistral AI（法国）Mistral AI (France)当前代current Small 4 / Medium 3.5Small 4 / Medium 3.5 · 2026-03
欧洲唯一全尺寸开放权重厂商。特点：开源杯（Small）与旗舰杯（Medium）许可不同——Small 4 是 Apache-2.0，Medium 3.5 是改版 MIT；下载前按档位读许可。Europe's only full-spectrum open-weight vendor. The open tier (Small) and flagship tier (Medium) carry different licences — Small 4 is Apache-2.0, Medium 3.5 is Modified MIT. Read the licence per tier before downloading.
未收量化对比，卡内说明原因 →No quantisation coverage — see the card →
MuseMuse
MetaMeta当前代current Muse GlimmerMuse Glimmer · 2026-08
Meta 在 Llama 之外的新开放权重线（meta-models org）。发布走门控：要先在 HF 登录同意才能下载，这是选型时的硬约束，卡上如实写。本站收录 Muse Glimmer 30B 的量化对比。Meta's newer open-weight line beside Llama (the meta-models org). Publication is gated: sign in and consent on HF before download — a hard constraint when choosing, stated plainly on the card. This site covers Muse Glimmer 30B quantisations.
看量化对比（Muse Glimmer）→Quantisation coverage (Muse Glimmer) →
OrnithOrnith
Ornith AI（DeepReinforce）Ornith AI (DeepReinforce)当前代current Ornith 1.5Ornith 1.5 · 2026-08
新晋家族（2026 年才出现两代）：MIT 许可 + 三档尺寸覆盖单卡到旗舰。出品主体 DeepReinforce 的注册辖区官方未声明，本站按「国外出品」处理并在属地依据里写明（S52 口径）。A newcomer (only two generations so far, both 2026): MIT licence plus three tiers from single-card to flagship. The publisher DeepReinforce's jurisdiction of registration is not officially stated; this site treats it as foreign-made and says so in the origin note (S52 convention).
看量化对比（Ornith 1.5）→Quantisation coverage (Ornith 1.5) →
PhiPhi
微软（Microsoft）Microsoft当前代current Phi-4-reasoning 系Phi-4-reasoning line · 2026-08 派生自 microsoft/Phi-4-reasoningderived from microsoft/Phi-4-reasoning
微软的小尺寸高质量路线：MIT、尺寸集中在 4B–15B、官方主打「合成数据教小模型」。当前代是 reasoning-vision 线（谱系上仍是 phi-4 后训练）。本站未收其量化对比——当前代 GGUF 生态太薄（见下）。Microsoft's small-but-high-quality line: MIT, sizes concentrated at 4B–15B, and the vendor's pitch is "teach small models with synthetic data". The current line is reasoning-vision (still a phi-4 post-train on the weight lineage). This site carries no quantisation comparison — the current generation's GGUF ecosystem is too thin (see below).
未收量化对比，卡内说明原因 →No quantisation coverage — see the card →
千问 QwenQwen
阿里巴巴 · 通义千问Alibaba · Qwen当前代current Qwen3.8Qwen3.8 · 2026-08
中文与代码口碑最强的开放权重家族之一；Apache-2.0 主力规格许可干净。本站量化对比的基准系列（qwen3-8）就在这一代。One of the strongest open-weight families for Chinese and code, with clean Apache-2.0 licensing on mainline builds. This site's benchmark quantisation series (qwen3-8) sits in this generation.
看量化对比（千问 3.8）→Quantisation coverage (Qwen3.8) →
腾讯混元Tencent Hunyuan
腾讯Tencent当前代current Hy4 previewHy4 preview · 2026-08
中文开放权重家族里少有的「大总参 + 高激活」路线；Hy4 把上下文拉到 1M+，对长程 agent 与工程代码有明确指向。本站暂无 GGUF 量化收录，选型前先看官方许可与硬件门槛。An uncommon large-total/high-active open-weight Chinese family; Hy4 pushes context past 1M and targets long-horizon agents and engineering code. This site has no GGUF coverage yet — read the official licence and hardware floor before choosing.
未收量化对比，卡内说明原因 →No quantisation coverage — see the card →
闭源谱系卡Closed-model cards 对照面：官方规格快照 · 每条挂出处 · 带复核到期日the contrast set: official spec snapshots · every fact sourced · with review due dates
本站主体是开放权重选型，但选型绕不开闭源 API 对照。每张闭源卡只记三件事：现在哪一代、官方声明过什么规格（每条挂官方出处）、我们上次核验是什么时候。闭源厂商改规格不通知任何人——所以每张卡带复核周期，到期会在页顶挂横幅，不装作永远最新。不采价格数字（D9），定价看官方定价页链接。This site is about open-weight selection, but that choice always sits next to the closed API alternative. Each closed card records exactly three things: which generation is current, what specs the vendor officially claims (every fact linked to its official source), and when we last verified it. Closed vendors change specs without notice — so every card carries a review cycle, and an overdue card shows a banner at the top instead of pretending to be current. No price figures (rule D9); pricing lives at the linked official pricing page.
Claude（闭源 API）Claude (closed API)
AnthropicAnthropic当前代current Claude 5.5 代（Fable 5.1 / Opus 5.5 / Sonnet 5.5 / Haiku 5.5）Claude 5.5 generation (Fable 5.1 / Opus 5.5 / Sonnet 5.5 / Haiku 5.5) · 2026-09
闭源旗舰线（agentic 编码与企业向）：官方模型总览是规格真相源；官方文档将模型 ID 定义为固定版本，而非持续更新的别名；选型时应核对具体 ID。A closed flagship line (agentic coding and enterprise): the official model overview is the spec source of truth; model IDs are pinned snapshots, not evergreen pointers (stated in official docs) — match versions by ID.
看官方规格快照 →Official spec snapshot →
豆包（闭源 API）Doubao (closed API)
字节跳动 · 火山引擎ByteDance · Volcengine当前代current Doubao-Seed 2.1 系（+ Seed-Evolving）Doubao-Seed 2.1 line (+ Seed-Evolving) · 2026-06
国产闭源主力：对话产品「豆包」与 API「Doubao-Seed」同族两条线；服务在网信办备案（Beijing-YunQue-20230821，豆包大模型=云雀更名，官方备案公示页确认）。代次更替快，本站卡带核验日与复核周期。The domestic closed flagship: the Doubao consumer app and the Doubao-Seed API are two lines of one family; the service is CAC-filed (Beijing-YunQue-20230821 — Doubao LLM is the renamed YunQue, confirmed on the official filing-disclosure page). Generations churn fast, so the card carries a check date and review interval.
看官方规格快照 →Official spec snapshot →
Gemini（闭源 API）Gemini (closed API)
Google DeepMindGoogle DeepMind当前代current Gemini 3 系（最新 3.8 Flash；3.1 Pro / 3.7 / 3.6 / 3.5 Flash 在册）Gemini 3 line (latest 3.8 Flash; 3.1 Pro / 3.7 / 3.6 / 3.5 Flash listed) · 2026-09
Google 的闭源旗舰线；与开放权重的 Gemma 是两条线，别混。选型注意官方「预览版」标注——预览意味着规格与价格都可能变，本站只抄快照。Google's closed flagship line; distinct from the open-weight Gemma line — do not conflate them. Note the official "preview" labels: preview means specs and prices may change; this site only snapshots.
看官方规格快照 →Official spec snapshot →
GPT（闭源 API）GPT (closed API)
OpenAIOpenAI当前代current GPT-6 系（Astra / 6.1 Sol / 6 Luna）GPT-6 line (Astra / 6.1 Sol / 6 Luna) · 2026-10
闭源旗舰线：能力声明以官方发布页为准，本站只抄规格不背书横评。注意 OpenAI 另有一条开放权重线（gpt-oss），与这条 API 线是两回事——本站两边都有卡。The closed flagship line: capability claims follow the official release pages; this site copies specs without endorsing comparisons. Note OpenAI also runs an open-weight line (gpt-oss) — a different thing from this API line; this site cards both.
看官方规格快照 →Official spec snapshot →
Grok（闭源 API）Grok (closed API)
xAI（官方文档页 2026-10 自称 SpaceXSI）xAI (official docs self-brand as SpaceXSI, Oct 2026)当前代current Grok 4.7Grok 4.7 · 2026-08
闭源旗舰线（代码与 agentic 向）：官方文档规格直抄。诚实标注：2026-10-20 快照里官方文档页自称 SpaceXSI（Memphis 数据中心口径），品牌归属变化本站不推测、只记录页面原文。A closed flagship line (code and agentic focus): specs copied from official docs. Honest note: in the 2026-10-20 snapshot the official docs self-brand as SpaceXSI (Memphis data-center wording); this site records the page text without speculating on brand changes.
看官方规格快照 →Official spec snapshot →
Kimi（闭源 API）Kimi (closed API)
月之暗面（Moonshot AI）Moonshot AI当前代current Kimi K3（API）Kimi K3 (API) · 2026-06
国产闭源旗舰，但它是本站唯一「API 与开放权重双轨」的家族：K3 权重在 Hugging Face / ModelScope 开放（自定义 Kimi K3 License），API 走 platform.kimi.com。选型时两条路径的许可与数据条款不同，分开读。The domestic closed flagship — yet the only family here with dual tracks: K3 weights are open on Hugging Face / ModelScope (custom Kimi K3 License) while the API runs on platform.kimi.com. Licences and data terms differ between the two paths — read them separately.
看官方规格快照 →Official spec snapshot →
做模型专题How models are made 从业者视角：训练四阶段 · 数据与算力 · 跑起来practitioner view: the four training stages · data & compute · serving
这一栏回答「模型是怎么被造出来、又是怎么跑起来的」。它与学 AI 站互链不重复：学 AI 教使用者视角的一页科普，这里是从业者视角的深版——讲决定、讲代价、讲口径变化。关键事实全部挂论文或官方模型卡出处；本站不实测、不裁决路线之争。This section answers "how models are built and how they run". It cross-links with the Learn site without duplicating it: Learn gives user-perspective one-page explainers; this is the practitioner-depth version — decisions, costs, and how the consensus changed. Key facts carry paper or official model-card sources; the site runs no benchmarks and adjudicates no route disputes.
预训练：模型怎么做到「什么都知道一点」Pre-training: how a model comes to know a bit of everything
做模型专题how models are made训练管线 · 第 1 站pipeline · stage 1
预训练在优化什么、数据怎么攒到几万亿 token、算力与数据怎么配比、缩放律改过几次口径。从业者视角的深版；使用者视角的一页科普在学 AI 站，两站互链不重复。What pre-training optimises, how corpora reach trillions of tokens, how compute and data are budgeted, and how the scaling-law story changed. The practitioner-depth version; the one-page explainer lives on the Learn site — cross-linked, not duplicated.
读这一篇 →Read this guide →
SFT：让续写机器学会「好好回答」SFT: teaching a completion machine to answer well
做模型专题how models are made训练管线 · 第 2 站pipeline · stage 2
监督微调在预训练之后做什么——指令数据从哪来、质量为什么压过数量、LoRA 这类省显存做法的边界。从业者视角深版；一页科普在学 AI 站。What supervised fine-tuning adds after pre-training — where instruction data comes from, why quality beats quantity, and the limits of memory-savers like LoRA. Practitioner depth; the one-page explainer lives on the Learn site.
读这一篇 →Read this guide →
偏好对齐：RLHF 与 DPO 在解决什么问题Preference alignment: what RLHF and DPO actually solve
做模型专题how models are made训练管线 · 第 3 站pipeline · stage 3
奖励模型怎么从人排序里炼出来、PPO 那一步贵在哪、DPO 为什么能免掉强化学习、以及「对齐税」与谄媚这些代价。从业者视角深版；一页科普在学 AI 站。How reward models are distilled from human rankings, why the PPO step is expensive, why DPO skips reinforcement learning, and the costs — alignment tax and sycophancy. Practitioner depth; the one-page explainer lives on the Learn site.
读这一篇 →Read this guide →
蒸馏：小模型怎么从大模型身上学Distillation: how small models learn from big ones
做模型专题how models are made训练管线 · 第 4 站pipeline · stage 4
logit 蒸馏与数据蒸馏两条路、为什么开放权重家族几乎人人蒸馏、以及「蒸馏来的」这个标签在选型里意味着什么。从业者视角深版；一页科普在学 AI 站。Logit distillation versus data distillation, why nearly every open-weight family distils, and what the "distilled" label means when choosing a model. Practitioner depth; the one-page explainer lives on the Learn site.
读这一篇 →Read this guide →
数据与算力：做模型的两大开销Data and compute: the two big bills of making a model
做模型专题how models are made训练管线 · 第 5 站pipeline · stage 5
语料从哪来、怎么洗到万亿 token；算力怎么用 GPU-hours 与 MFU 记账；成本量级怎么估——只讲构成方法，不存任何单价。从业者视角；词条版在学 AI 站，互链不重复。Where corpora come from and how they get washed down to trillions of tokens; how compute is booked in GPU-hours and MFU; how to estimate cost magnitude — method only, no unit prices stored. Practitioner view; the term pages live on the Learn site, cross-linked, not duplicated.
读这一篇 →Read this guide →
开源训练栈：多卡训练的工具链The open training stack: the toolchain behind multi-GPU training
做模型专题how models are made训练管线 · 第 6 站pipeline · stage 6
一张卡装不下大模型之后，数据并行、ZeRO、张量并行、流水线并行怎么组合；DeepSpeed、Megatron-LM、torchtitan、nanotron、TRL 各站在哪一层。从业者视角的选型地图；词条版在学 AI 站，互链不重复。Once a model stops fitting on one GPU: how data parallelism, ZeRO, tensor and pipeline parallelism compose, and which layer DeepSpeed, Megatron-LM, torchtitan, nanotron and TRL each sit at. A practitioner's selection map; term pages live on the Learn site, cross-linked, not duplicated.
读这一篇 →Read this guide →
跑起来深水：KV 缓存、批处理与推理引擎Serving deep dive: KV cache, batching, and inference engines
做模型专题how models are made训练管线 · 第 7 站pipeline · stage 7
推理吞吐的瓶颈为什么在 KV 缓存显存；PagedAttention、连续批处理、RadixAttention 各解决什么；vLLM 与 SGLang 站在哪。从业者视角；一页科普在学 AI 站，互链不重复。Why serving throughput is bottlenecked by KV-cache memory; what PagedAttention, continuous batching and RadixAttention each solve; where vLLM and SGLang sit. Practitioner view; the one-page explainers live on the Learn site, cross-linked, not duplicated.
读这一篇 →Read this guide →
GPU 事实与成本数学：跑模型的硬件账GPU facts and cost math: the hardware bill for running a model
做模型专题how models are made训练管线 · 第 8 站pipeline · stage 8
从可用内存拆解驻留权重、上下文缓存、并发与运行时；区分理想化带宽算式和实际吞吐，附可填写预算，不存单价或购卡建议。Separate resident weights, context cache, concurrency and runtime; distinguish idealized bandwidth arithmetic from actual throughput, with an editable budget and no stored prices or purchase advice.
读这一篇 →Read this guide →
任务模型：嵌入、重排、OCR 与语音识别Task models: embedding, reranking, OCR and ASR
做模型专题how models are made训练管线 · 第 9 站pipeline · stage 9
从输入输出、运行库和验收方法区分四类任务；逐层记录平台与格式兼容性，未安装验证的组合保留未知，附练习与排查入口。Distinguish four task contracts by input, output, runtime and acceptance; record platform/format compatibility by layer, retain untested combinations as unknown, and follow practice and troubleshooting links.
读这一篇 →Read this guide →
把日志留在本地到底意味着什么What keeping logs on your own machine actually means
做模型专题how models are made把「日志留在本地」拆成三条线：写在哪、谁能读、留多久。逐条给运行器官方文档里的证据位置，并说明磁盘加密、备份同步与监听地址为什么都不在这四个字里。Unpack "logs stay local" into three lines: where they are written, who can read them, and how long they persist. Each item points at the runner's own documentation, and explains why disk encryption, backup sync and bind address are not contained in those four words.
读这一篇 →Read this guide →
多模态模型的显存账怎么算Counting memory for a multimodal model
做模型专题how models are made把多模态的占用分成编码器、投影件与解码器缓存三笔账，写清每笔的来源、单位（GiB）与算法边界；KV 算式只管同层标准全注意力解码器，不能外推到编码器与 MLA、混合、滑动窗口。Split multimodal occupancy into three bills — encoder, projector files and decoder cache — with the source, unit (GiB) and scope stated for each. The KV formula covers only a standard same-layer full-attention decoder; it does not extend to encoders or to MLA, hybrid and sliding-window designs.
读这一篇 →Read this guide →
本地跑起来之前先确认的七件事Seven checks before a local run starts
做模型专题how models are made从开机到第一次出字的七个核对点：可用容量、驻留体积、文件与运行器、缓存口径、许可、结果验收、日志落点。每件写清查什么、拿什么当证据、失败后下一步。Seven checkpoints between power-on and the first token: available capacity, resident size, file and runner support, cache scope, licence, output acceptance and where logs go. Each states what to check, what counts as evidence and the next step on failure.
读这一篇 →Read this guide →
量化档位与精度损失：能说什么、不能说什么Quantisation tiers and accuracy loss: what can be said, and what cannot
做模型专题how models are made把档位名、每权重比特数与精度损失分成三层：哪些是发布页事实、哪些是本站按命名标准的推导、哪些必须留空；顺带写清换档后容量账里哪一笔会变。Split tier names, bits per weight and accuracy loss into three layers: publishing-page facts, this site's derivation from the public naming standard, and what must stay blank. It also states which line in the memory bill actually moves when a tier changes.
读这一篇 →Read this guide →
怎么按官方口径读一个模型的评测说明How to read a model's evaluation write-up on its official terms
做模型专题how models are made读一份官方评测说明要补齐的四件事：任务集范围、指标口径、对照与版本、声明出处。哪些话有依据、哪些话只能留空，以及自报数字与你的任务之间隔着什么。The four things to fill in when reading an official evaluation write-up: task-set scope, metric definition, comparison and version, and where the claim comes from. What can be stated, what must stay blank, and what separates a self-reported number from your task.
读这一篇 →Read this guide →
怎么用这个站How to use this site
- 先看你有多少显存。这是唯一的硬约束，8GB 和 24GB 能跑的档位完全不同。显存撑不下时看同一张表的「系统内存 / 多卡」两栏——纯 CPU 或多卡张量并行也是可行路线。Start with how much VRAM you have. It is the only hard constraint — the tiers that fit in 8GB and 24GB are entirely different. If VRAM is the limit, look at the "system memory / multi-GPU" columns in the same table — CPU-only or multi-GPU tensor parallel are viable routes.
- 在 B 区块初筛文件体积候选。再核对缓存、并发和运行时；表格不保证能运行。Use block B to filter file-size candidates. Then budget cache, concurrency and runtime; it is not an execution guarantee.
- 在 C 区块横向对比量化者。看谁档位全、谁带 imatrix、谁在跟进更新。Compare quantisers side by side in block C. See who covers the most tiers, who ships imatrix, and who is keeping up with releases.
- 在 D 区块看单档详解。七个字段说清每个变体是什么、差在哪、什么时候该选它。Read the per-tier detail in block D. Seven fields spell out what each variant is, how it differs, and when to pick it.
- 下载前先读许可。非宽松许可的，页面会标出来；量化仓没单独声明许可的，页面会写明「继承自基座」，原文链接就在旁边。Read the licence before downloading. Non-permissive licences are flagged; where a quant repo declares none, the page says "inherited from base" with the original link right beside it.
已收录系列Series covered 第二层：同一模型该下哪个量化版layer two: which quantisation of the same model · 按 GGUF 量化实际下载量排序ranked by measured GGUF download volume
DeepSeek V4DeepSeek V4
DeepSeekDeepSeekMIT 许可，量化生态活跃，antirez（Redis 作者、llama.cpp 生态参与者）维护量化版。MIT licensed with an active quantisation ecosystem — the quantisations are maintained by antirez (author of Redis, a participant in the llama.cpp ecosystem) himself.
- 通用对话General chat
- MIT 许可MIT licence
- 推理Reasoning
看该选哪一档 →Which tier to pick →
Gemma 4Gemma 4
GoogleGoogleGoogle 开放权重系列，本站收录的三种规格各自都有量化覆盖。Google's open-weight series; the three sizes listed here each have quantisation coverage.
- 通用对话General chat
- MoE 稀疏Sparse MoE
- 小尺寸可跑Runnable at small sizes
看该选哪一档 →Which tier to pick →
GLM 5GLM 5
智谱 AIZhipu AI智谱开放权重系列；不同规格的许可分别标注，部署前核对对应版本与容量需求。Zhipu's open-weight series. Licences are listed per variant; check the specific version and capacity requirements before deployment.
- 通用对话General chat
- 工具调用Tool use
- MIT 许可MIT licence
看该选哪一档 →Which tier to pick →
gpt-ossgpt-oss
OpenAIOpenAIOpenAI 开放权重系列，本站索引 20B 规格的量化信息；是否适合单卡取决于可用内存、量化和上下文配置。OpenAI's open-weight series, with quantisation information for the 20B variant. Single-GPU suitability depends on available memory, quantisation and context settings.
- 推理Reasoning
- MoE 稀疏Sparse MoE
- apache-2.0Apache-2.0
看该选哪一档 →Which tier to pick →
KAT Coder V2.5KAT Coder V2.5
KwaipilotKwaipilot面向代码的开放权重系列，apache-2.0，34 个第三方量化仓。An open-weight series aimed at code, apache-2.0, with 34 third-party quantisation repos.
- 工具调用Tool use
- 长上下文Long context
- apache-2.0Apache-2.0
看该选哪一档 →Which tier to pick →
Kimi K3Kimi K3
月之暗面Moonshot AI月之暗面开放权重系列，快照收录 41 家量化仓；数量不代表模型质量或运行兼容性。Moonshot's open-weight series, with 41 quantisation repos in the snapshot. Counts do not establish quality or runtime compatibility.
- 通用对话General chat
- 长上下文Long context
看该选哪一档 →Which tier to pick →
Laguna 2.1Laguna 2.1
PoolsidePoolside代码方向的开放权重系列，许可为 OpenMDW-1.1（使用前须读原文）。An open-weight series in the coding direction; licensed OpenMDW-1.1 (read the original terms before use).
- 工具调用Tool use
- 长上下文Long context
看该选哪一档 →Which tier to pick →
MiniCPM 5MiniCPM 5
面壁智能 · OpenBMBOpenBMB国产小尺寸代表，2B 规模适合低显存与端侧。A representative Chinese small-size model; the 2B scale suits low VRAM and on-device use.
- 端侧On-device
- 小尺寸Small sizes
- 中文强项Strong Chinese
看该选哪一档 →Which tier to pick →
Muse GlimmerMuse Glimmer
MetaMetaMeta 的开放权重系列，apache-2.0，量化覆盖横跨多家主力量化者。Meta's open-weight series, apache-2.0, with quantisation coverage across several major quantisers.
- 通用对话General chat
- apache-2.0Apache-2.0
看该选哪一档 →Which tier to pick →
Ornith 1.5Ornith 1.5
Ornith AIOrnith AI模型卡标注 MIT 许可，两个规格都有 GGUF 量化覆盖；具体文件与版本见下表。The model cards list MIT; both sizes have GGUF coverage. See the tables for individual files and versions.
- 通用对话General chat
- MIT 许可MIT licence
- MoE 稀疏Sparse MoE
看该选哪一档 →Which tier to pick →
千问 3.8Qwen3.8
阿里 · 通义千问Alibaba Qwen阿里千问系列，本站收录 27B 密集模型与 Flash-Next 的量化信息；收录不代表性能排名。Alibaba's Qwen series, with quantisation information for the dense 27B model and Flash-Next. Inclusion is not a performance ranking.
- 通用对话General chat
- 工具调用Tool use
- 长上下文Long context
- 中文强项Strong Chinese
看该选哪一档 →Which tier to pick →
收录的量化者Quantisers covered 54 位of them
6block
Abiray
AesSedai
Amynnnn
AtomicChat
Coder40-95
DhruvalLabs
GrEarl
JigSawPT
KikoCis
Kremon96
Kuberwastaken
LemesoftNostalgic
Mungert
NANI-Nithin
Ninnix96
RemySkye
SandLogicTechnologies
Solstice-AI
TheStageAI
WhiskyAKM
abenzerps
aj9o9
apetersson
aria220
bartowski第三方量化发布者，档位与方法以仓库说明为准Third-party quantiser; check repository notes for tiers and methods
batiai
byteshape
ddh0
dinerburger
eaddario
ggml-orgllama.cpp 所属组织的官方量化发布Official quant builds from the llama.cpp org
gguf
giladgd
hudabey
kizzlah
lausannequants
lmstudio-communityLM Studio 官方社区组织，为其 GUI 客户端供包Powers the LM Studio GUI
meta-models
mradermacher发布低比特量化，具体方法与来源见仓库说明Publishes low-bit quants; check repository methods and provenance
mxxm-t
na0x2c6
ngquocvinh
oki0ki
openbmb模型方发布的 GGUF，仍需核对版本与运行器支持Published by the model authors themselves
ornith-ai
poolside
prithivMLmods
qtum
robertxxiv
smalinin
unsloth第三方量化发布者，具体覆盖与下载量见快照Third-party quantiser; see snapshot coverage and downloads
vcruz305
vultr
数据来源与法律立场Sources and legal position
数据来自Data comes from the Hugging Face 公开 API（模型元数据 + 文件体积），采集public API (model metadata + file sizes), collected时间at 2026-10-30，为快照数据而非实时。; this is a snapshot, not live data.
本站不跑 benchmark，不给质量与速度结论。凡涉及主观判断处均标注「我们的口径」。This site does not run benchmarks and does not draw quality or speed conclusions. Every place we offer a judgement is marked as our own reading.
不下载、不托管任何权重文件No weight files are downloaded or hosted here；许可原样引自发布页，本站不解读许可、不构成法律意见，一切以; licences are quoted verbatim from release pages, this site does not interpret them or give legal advice, and everything is governed by发布页原文the legal page为准。.
系列选择依据四维打分（量化热度 40 / 量化广度 25 / 规模可及性 20 / 新鲜度 15），方法与落选名单见 Series selection uses a four-factor score (quantisation heat 40 / coverage 25 / size reachability 20 / freshness 15); the method and the rejected candidates are documented in 数据仓 METHODOLOGY.md ↗the data repo's METHODOLOGY.md ↗。.