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
先说清楚：这个站的数据怎么来的First, where this site's data comes from
✅ 实测字段Measured fields
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.
✅ 推导Derived
由实测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 measured fields by published rules: bits per weight (per the llama.cpp naming convention), tier assignment, VRAM requirement range. The rules are fully published inMETHODOLOGY.md 完整公开，任何人可复算。 — anyone can recompute them.
❌ 我们不说的What we do not claim
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)
出品属地Origin国产domestic出品方 Kwaipilot 是快手（北京）的 AI 团队——官方 HF 组织页自述「the AI explorers from Kuaishou」。Kwaipilot is the AI team of Kuaishou (Beijing) — its official HF org page says "the AI explorers from Kuaishou".
国内下载Direct access国内可直连direct OKModelScope 官方镜像直连可达（2026-10-20 实测 200）；Hugging Face 直连不可达。Official ModelScope mirror reachable (measured 200); Hugging Face direct unreachable. · 国内镜像入口domestic mirror
备案状态CAC filing未见于备案名单not in filing list：「KAT Coder」未出现在网信办公示名单（2026-10-20 检索）；快手旗下「快意大模型」「可图大模型」「可灵」等已备案，但本站不替 KAT Coder 推断。“KAT Coder” does not appear in the CAC public registry (searched 2026-10-20); Kuaishou’s “Kuaiyi”, “Ketu” and “Kling” models are filed, but this site does not infer a filing for KAT Coder.
备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「国内可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise.
网信办《生成式人工智能服务已备案信息》滚动公告页（13 个批次附件 PDF，2026-10-20 全量下载检索）：CAC · filed generative-AI service information (rolling announcement page, 13 batch attachment PDFs, full download & search on 2026-10-20): https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm
Kwaipilot 官方 Hugging Face 组织页（自述来自快手）：Kwaipilot official Hugging Face org page (self-described as from Kuaishou): https://huggingface.co/Kwaipilot
✅ 公开下载 无需申请✅ Public download no application needed
规模Scale
官网标注Vendor states
34.66B 总参34.66B total params
官方 API 申报Official API
—
本站实测反推This site, measured
—
权重精度构成Precision breakdown
—
架构Architecture
qwen35moe
上下文长度Context length
262,144 tokens
基座Base model
本仓即基座（不是微调版）this repo is the base model, not a fine-tune
任务类型Task type
—
输入模态Input modality
✗ 不支持✗ Not supported出处 ↗source ↗模型卡model card · 2026-10-28 模型卡明确：开放权重版仅语言模型权重、纯文本；视觉组件未包含（HF 标签里的 image-text-to-text 与卡面说明冲突，以卡面为准）。Card states the open-weight release ships language-model weights only and operates as a text-only model; vision components are not included (the HF image-text-to-text tag conflicts with the card; the card wins).
apache-2.0，34 个第三方量化仓（mudler 805k（APEX 仓，原 Dev-GGUF 仓已下架）/ bartowski 279k）。基准仓自身下载量不高，但量化采用集中。apache-2.0, 34 third-party quantisation repos (mudler 805k on the APEX repo; the original Dev-GGUF repo has been taken down / bartowski 279k). The base repo itself has modest downloads, but quantisation uptake is concentrated.
B该选哪个Which one to pick按你的显存倒推work backwards from your VRAM
按「你实际能腾出多少显存或内存」选档位，而不是按参数猜。下面每档的体积都是实测值（来自发布页文件列表），加权 15% 余量给运行时与 KV cache。Pick a tier by how much VRAM or memory you can actually free up, not by guessing from spec sheets. Every size below is measured (from the release page file listing) with 15% headroom added for runtime and the KV cache.
单卡全量装得下Fits entirely on one card
显存 → 档位速查（权重全部放显存，不卸载到系统内存）VRAM → tier cheat sheet (all weights on the GPU, no offload to system memory)
8 GB 显存8 GB VRAM
显存装不下最小档VRAM cannot hold even the smallest tier（最小 9.8 GB） (smallest tier 9.8 GB) 共享显卡 / 老卡，够跑小MoE 与 2B 级shared or older cards; enough for small MoE and 2B-class models
Q8_0Q8_0（37 GB · 近似无损Near lossless）
更省：Smaller: Q6_KQ6_K 29 GB / Q5_1Q5_1 26 GB A100 80G / H100 80G / Mac Studio 大统一内存A100 80G / H100 80G / Mac Studio large unified memory
显存不够时的替代路径When VRAM is not enough
llama.cpp 的llama.cpp’s-ngl可以只把部分层放显存、其余落系统内存；多卡则自动切分。can keep only some layers on the GPU and spill the rest to system memory; with multiple cards it splits automatically.代价是速度：纯 CPU 推理通常比同规模 GPU 慢数倍，但同样能出结果。The trade-off is speed: CPU-only inference is usually several times slower than a comparable GPU, but still produces results.
Q8_0Q8_0（37 GB · 近似无损Near lossless）
更省：Smaller: Q6_KQ6_K 29 GB / Q5_1Q5_1 26 GB 纯 CPU 推理，速度慢但可跑；建议开swap 兜底CPU-only inference — slow but it runs; enable swap as a safety net 纯 CPU 会明显慢于 GPU；若显存够，优先用上面的单卡档 CPU-only is markedly slower than GPU — if you have the VRAM, prefer the single-card tiers above
128 GB 系统内存128 GB system RAM
Q8_0Q8_0（37 GB · 近似无损Near lossless）
更省：Smaller: Q6_KQ6_K 29 GB / Q5_1Q5_1 26 GB 工作站标准配置，纯 CPU 可跑中等规模 MoEthe standard workstation config; CPU-only handles mid-size MoE 纯 CPU 会明显慢于 GPU；若显存够，优先用上面的单卡档 CPU-only is markedly slower than GPU — if you have the VRAM, prefer the single-card tiers above
256 GB 统一内存 / 多卡256 GB unified memory / multi-GPU
Q8_0Q8_0（37 GB · 近似无损Near lossless）
更省：Smaller: Q6_KQ6_K 29 GB / Q5_1Q5_1 26 GB Mac Studio / 服务器内存，或 2~4 张 80G 卡张量并行Mac Studio / server memory, or 2–4 × 80G cards under tensor parallel 纯 CPU 会明显慢于 GPU；若显存够，优先用上面的单卡档 CPU-only is markedly slower than GPU — if you have the VRAM, prefer the single-card tiers above
640 GB+ 内存 / 8×80G640 GB+ memory / 8×80G
Q8_0Q8_0（37 GB · 近似无损Near lossless）
更省：Smaller: Q6_KQ6_K 29 GB / Q5_1Q5_1 26 GB 超大 MoE 的门槛，需要多卡切分或大内存服务器the threshold for very large MoE — needs multi-GPU splitting or a big-memory server 纯 CPU 会明显慢于 GPU；若显存够，优先用上面的单卡档 CPU-only is markedly slower than GPU — if you have the VRAM, prefer the single-card tiers above
本系列档位体积跨度 9.8 GB → 69 GB。
Tier sizes in this series span 9.8 GB → 69 GB.
跨度过大说明同一系列的量化选择空间很宽，这正是量化选型的主要收益所在。
展开：本系列所有档位的实测体积Expand: measured sizes for every tier（28 档tiers）
量化档Quant
最小体积Min size
每权重比特bpw
档位Tier
这档意味着什么What this tier means
该档文件数 / 体积跨度Files in tier / size range
最小体积来源
IQ2_XXSIQ2_XXS
9.8 GB
2.13 bpw
极限压缩Extreme compression
能装下就装，质量明显下降Fit it in at the cost of visibly lower quality
「每权重比特」是按 llama.cpp 公开命名标准从文件名推导的，同一档位在不同模型上的真实值会略有差异。这是Bits per weight is derived from the filename by llama.cpp’s published naming convention; the real value varies slightly between models at the same tier. This isB 级推导grade-B derived，不是实测值。 — not a measured value.
卡片左侧色条 = 推荐档位Colour bar on the left = recommended tier近似无损Near lossless保守Conservative平衡Balanced长上下文Long context极限压缩Extreme compression
「建议档」是本站按体积最小且落在平衡档的规则自动选出的，不是质量排名。官方量化排在最前；带 imatrix 标记的仓库带重要性矩阵数据。The recommended tier is picked by rule — smallest size that still lands in the balanced tier — not by a quality ranking. Official quantisations are listed first; repos marked imatrix ship an importance matrix.
D每个变体详解Every variant in detail七字段seven fields
每个变体按七个字段展开：它是什么 → 量化者做了什么 → 与同系列其他变体的区别 → 存在的意义 → 优势 → 为什么选它 → 已知取舍。最后一项是本站唯一会写主观判断的地方，且都标注了「我们的口径」。Each variant expands along seven fields: what it is → what the quantiser did → how it differs from other variants in this series → why it exists → strengths → why you would pick it → known trade-offs.The last field is the only place this site offers a subjective judgement, and every item is marked as our own reading.
量化者做了什么What the quantiser did带重要性矩阵（imatrix：/models_out/KAT-Coder-V2.5-Dev-GGUF/Kwaipilot_KAT-Coder-V2.5-Dev-imatrix.gguf），用激活数据统计各层重要性后再量化，通常比纯 RTN 稳Ships an importance matrix (imatrix: /models_out/KAT-Coder-V2.5-Dev-GGUF/Kwaipilot_KAT-Coder-V2.5-Dev-imatrix.gguf) — quantises after scoring each layer by activation statistics, usually steadier than plain RTN；针对 qwen35moe 架构优化Optimised for the qwen35moe architecture；声明上下文 262,144 tokensDeclared context 262,144 tokens；发布页声明溯源到 Kwaipilot/KAT-Coder-V2.5-DevThe release page traces this back to Kwaipilot/KAT-Coder-V2.5-Dev；权重总字节 568.8 GB（发布页 gguf.total 字段）Total weight bytes 568.8 GB (the release page's gguf.total field)。
与同系列其他变体的区别How it differs from other variants in this series同系列另有 5 个量化来源（mradermacher 11 档tiers、RemySkye 16 档tiers、robertxxiv 3 档tiers、Ninnix96 0 档tiers 等）。The series has 5 other quantisation sources (mradermacher 11 tiers, RemySkye 16 tiers, robertxxiv 3 tiers, Ninnix96 0 tiers, among others). 本仓 26 档，覆盖最全。 This repo has 26 tiers — the broadest coverage coverage in this series.
存在的意义Why it exists补上官方没覆盖的档位（Q4_1 这类），让你能在有限显存里凑合跑起来，或者反过来用更高精度换质量。Fills tiers the official build does not cover (such as Q4_1), so you can get it running on limited VRAM — or trade the other way, for more precision and quality.
优势Strengths· 档位覆盖最广的一档（26 档），从极限压缩到近似无损都能选The widest tier coverage here (26 tiers), from extreme compression up to near lossless · 带 imatrix，同档位下质量通常更稳Ships imatrix, so quality at the same tier is usually steadier · 平衡档落在 Q4_1（22 GB），是大多数人的起点The balanced tier lands on Q4_1 (22 GB), the starting point for most people
为什么选它 / 目的Why you would pick it / its purpose如果你的瓶颈是显存：直接下 Q4_1（22 GB，平衡档）。 如果你的瓶颈是速度：这个仓里没有更快的档位（GGUF 的速度主要由架构与后端决定，不由量化档决定），换量化者意义不大。 如果你的瓶颈是质量：往上走一档到保守档（Q5_K_M / Q6_K）。If VRAM is your constraint: take Q4_1 directly (22 GB, Balanced). If speed is your constraint: this repo has nothing faster — GGUF speed is decided by architecture and backend, not by quantisation tier — so switching quantiser will not help. If quality is your constraint: step up one tier to conservative (Q5_K_M / Q6_K).
已知取舍Known trade-offs· 全仓 530 GB，硬盘紧张的话建议只下需要的 1~2 个档位530 GB for the whole repo — if disk is tight, download only the one or two tiers you need
量化者做了什么What the quantiser did未提供重要性矩阵（imatrix），走的是纯权重量化路线No importance matrix (imatrix); this is pure weight quantisation；针对 qwen35moe 架构优化Optimised for the qwen35moe architecture；声明上下文 262,144 tokensDeclared context 262,144 tokens；发布页声明溯源到 Kwaipilot/KAT-Coder-V2.5-DevThe release page traces this back to Kwaipilot/KAT-Coder-V2.5-Dev；权重总字节 237.1 GB（发布页 gguf.total 字段）Total weight bytes 237.1 GB (the release page's gguf.total field)。
与同系列其他变体的区别How it differs from other variants in this series同系列另有 5 个量化来源（bartowski 26 档tiers、RemySkye 16 档tiers、robertxxiv 3 档tiers、Ninnix96 0 档tiers 等）。The series has 5 other quantisation sources (bartowski 26 tiers, RemySkye 16 tiers, robertxxiv 3 tiers, Ninnix96 0 tiers, among others). 本仓 11 档，覆盖中等。 This repo has 11 tiers — middling coverage in this series.
存在的意义Why it exists补上官方没覆盖的档位（Q4_K_M 这类），让你能在有限显存里凑合跑起来，或者反过来用更高精度换质量。Fills tiers the official build does not cover (such as Q4_K_M), so you can get it running on limited VRAM — or trade the other way, for more precision and quality.
优势Strengths· 平衡档落在 Q4_K_M（21 GB），是大多数人的起点The balanced tier lands on Q4_K_M (21 GB), the starting point for most people
为什么选它 / 目的Why you would pick it / its purpose如果你的瓶颈是显存：直接下 Q4_K_M（21 GB，平衡档）。 如果你的瓶颈是速度：这个仓里没有更快的档位（GGUF 的速度主要由架构与后端决定，不由量化档决定），换量化者意义不大。 如果你的瓶颈是质量：往上走一档到保守档（Q5_K_M / Q6_K）。If VRAM is your constraint: take Q4_K_M directly (21 GB, Balanced). If speed is your constraint: this repo has nothing faster — GGUF speed is decided by architecture and backend, not by quantisation tier — so switching quantiser will not help. If quality is your constraint: step up one tier to conservative (Q5_K_M / Q6_K).
已知取舍Known trade-offs· 无 imatrix，同档位下质量损失可能比带 imatrix 的版本更明显（我们的口径，需实测验证）No imatrix, so quality loss at the same tier may be more visible than in an imatrix build (our reading, to be verified by measurement) · 全仓 221 GB，硬盘紧张的话建议只下需要的 1~2 个档位221 GB for the whole repo — if disk is tight, download only the one or two tiers you need
量化者做了什么What the quantiser did未提供重要性矩阵（imatrix），走的是纯权重量化路线No importance matrix (imatrix); this is pure weight quantisation；针对 qwen35moe 架构优化Optimised for the qwen35moe architecture；声明上下文 262,144 tokensDeclared context 262,144 tokens；发布页声明溯源到 Kwaipilot/KAT-Coder-V2.5-DevThe release page traces this back to Kwaipilot/KAT-Coder-V2.5-Dev；权重总字节 398.2 GB（发布页 gguf.total 字段）Total weight bytes 398.2 GB (the release page's gguf.total field)。
与同系列其他变体的区别How it differs from other variants in this series同系列另有 5 个量化来源（bartowski 26 档tiers、mradermacher 11 档tiers、robertxxiv 3 档tiers、Ninnix96 0 档tiers 等）。The series has 5 other quantisation sources (bartowski 26 tiers, mradermacher 11 tiers, robertxxiv 3 tiers, Ninnix96 0 tiers, among others). 本仓 16 档，覆盖中等。 This repo has 16 tiers — middling coverage in this series.
存在的意义Why it exists补上官方没覆盖的档位（Q4_1 这类），让你能在有限显存里凑合跑起来，或者反过来用更高精度换质量。Fills tiers the official build does not cover (such as Q4_1), so you can get it running on limited VRAM — or trade the other way, for more precision and quality.
优势Strengths· 平衡档落在 Q4_1（22 GB），是大多数人的起点The balanced tier lands on Q4_1 (22 GB), the starting point for most people
为什么选它 / 目的Why you would pick it / its purpose如果你的瓶颈是显存：直接下 Q4_1（22 GB，平衡档）。 如果你的瓶颈是速度：这个仓里没有更快的档位（GGUF 的速度主要由架构与后端决定，不由量化档决定），换量化者意义不大。 如果你的瓶颈是质量：往上走一档到保守档（Q5_K_M / Q6_K）。If VRAM is your constraint: take Q4_1 directly (22 GB, Balanced). If speed is your constraint: this repo has nothing faster — GGUF speed is decided by architecture and backend, not by quantisation tier — so switching quantiser will not help. If quality is your constraint: step up one tier to conservative (Q5_K_M / Q6_K).
已知取舍Known trade-offs· 无 imatrix，同档位下质量损失可能比带 imatrix 的版本更明显（我们的口径，需实测验证）No imatrix, so quality loss at the same tier may be more visible than in an imatrix build (our reading, to be verified by measurement) · 全仓 371 GB，硬盘紧张的话建议只下需要的 1~2 个档位371 GB for the whole repo — if disk is tight, download only the one or two tiers you need
量化者做了什么What the quantiser did未提供重要性矩阵（imatrix），走的是纯权重量化路线No importance matrix (imatrix); this is pure weight quantisation；针对 qwen35moe 架构优化Optimised for the qwen35moe architecture；声明上下文 262,144 tokensDeclared context 262,144 tokens；发布页声明溯源到 Qwen3.6-35B-A3BThe release page traces this back to Qwen3.6-35B-A3B；权重总字节 119.1 GB（发布页 gguf.total 字段）Total weight bytes 119.1 GB (the release page's gguf.total field)。
与同系列其他变体的区别How it differs from other variants in this series同系列另有 5 个量化来源（bartowski 26 档tiers、mradermacher 11 档tiers、RemySkye 16 档tiers、Ninnix96 0 档tiers 等）。The series has 5 other quantisation sources (bartowski 26 tiers, mradermacher 11 tiers, RemySkye 16 tiers, Ninnix96 0 tiers, among others). 本仓 3 档，覆盖中等。 This repo has 3 tiers — middling coverage in this series.
存在的意义Why it exists补上官方没覆盖的档位（Q4_K_M 这类），让你能在有限显存里凑合跑起来，或者反过来用更高精度换质量。Fills tiers the official build does not cover (such as Q4_K_M), so you can get it running on limited VRAM — or trade the other way, for more precision and quality.
优势Strengths· 平衡档落在 Q4_K_M（21 GB），是大多数人的起点The balanced tier lands on Q4_K_M (21 GB), the starting point for most people
为什么选它 / 目的Why you would pick it / its purpose如果你的瓶颈是显存：直接下 Q4_K_M（21 GB，平衡档）。 如果你的瓶颈是速度：这个仓里没有更快的档位（GGUF 的速度主要由架构与后端决定，不由量化档决定），换量化者意义不大。 如果你的瓶颈是质量：往上走一档到保守档（Q5_K_M / Q6_K）。If VRAM is your constraint: take Q4_K_M directly (21 GB, Balanced). If speed is your constraint: this repo has nothing faster — GGUF speed is decided by architecture and backend, not by quantisation tier — so switching quantiser will not help. If quality is your constraint: step up one tier to conservative (Q5_K_M / Q6_K).
已知取舍Known trade-offs· 档位极少，没有升降空间，被显存卡住时无路可退Very few tiers, so there is no room to move up or down when VRAM runs out · 无 imatrix，同档位下质量损失可能比带 imatrix 的版本更明显（我们的口径，需实测验证）No imatrix, so quality loss at the same tier may be more visible than in an imatrix build (our reading, to be verified by measurement)
量化者做了什么What the quantiser did带重要性矩阵（imatrix：gguf/Kwaipilot_KAT-Coder-V2.5-Dev-imatrix.gguf），用激活数据统计各层重要性后再量化，通常比纯 RTN 稳Ships an importance matrix (imatrix: gguf/Kwaipilot_KAT-Coder-V2.5-Dev-imatrix.gguf) — quantises after scoring each layer by activation statistics, usually steadier than plain RTN；针对 qwen35moe 架构优化Optimised for the qwen35moe architecture；声明上下文 262,144 tokensDeclared context 262,144 tokens；发布页声明溯源到 Kwaipilot/KAT-Coder-V2.5-DevThe release page traces this back to Kwaipilot/KAT-Coder-V2.5-Dev。
与同系列其他变体的区别How it differs from other variants in this series同系列另有 5 个量化来源（bartowski 26 档tiers、mradermacher 11 档tiers、RemySkye 16 档tiers、robertxxiv 3 档tiers 等）。The series has 5 other quantisation sources (bartowski 26 tiers, mradermacher 11 tiers, RemySkye 16 tiers, robertxxiv 3 tiers, among others). 本仓 0 档，覆盖最少。 This repo has 0 tiers — the narrowest coverage in this series.
存在的意义Why it exists补上官方没覆盖的档位（— 这类），让你能在有限显存里凑合跑起来，或者反过来用更高精度换质量。Fills tiers the official build does not cover (such as —), so you can get it running on limited VRAM — or trade the other way, for more precision and quality.
优势Strengths· 带 imatrix，同档位下质量通常更稳Ships imatrix, so quality at the same tier is usually steadier
为什么选它 / 目的Why you would pick it / its purpose该仓无可用档位。This repo has no usable tier.
已知取舍Known trade-offs· 档位极少，没有升降空间，被显存卡住时无路可退Very few tiers, so there is no room to move up or down when VRAM runs out
量化者做了什么What the quantiser did未提供重要性矩阵（imatrix），走的是纯权重量化路线No importance matrix (imatrix); this is pure weight quantisation；针对 qwen35moe 架构优化Optimised for the qwen35moe architecture；声明上下文 262,144 tokensDeclared context 262,144 tokens；发布页声明溯源到 Kwaipilot/KAT-Coder-V2.5-DevThe release page traces this back to Kwaipilot/KAT-Coder-V2.5-Dev；权重总字节 30.8 GB（发布页 gguf.total 字段）Total weight bytes 30.8 GB (the release page's gguf.total field)。
与同系列其他变体的区别How it differs from other variants in this series同系列另有 5 个量化来源（bartowski 26 档tiers、mradermacher 11 档tiers、RemySkye 16 档tiers、robertxxiv 3 档tiers 等）。The series has 5 other quantisation sources (bartowski 26 tiers, mradermacher 11 tiers, RemySkye 16 tiers, robertxxiv 3 tiers, among others). 本仓 1 档，覆盖中等。 This repo has 1 tiers — middling coverage in this series.
存在的意义Why it exists补上官方没覆盖的档位（Q6_K 这类），让你能在有限显存里凑合跑起来，或者反过来用更高精度换质量。Fills tiers the official build does not cover (such as Q6_K), so you can get it running on limited VRAM — or trade the other way, for more precision and quality.
优势Strengths· 平衡档落在 Q6_K（31 GB），是大多数人的起点The balanced tier lands on Q6_K (31 GB), the starting point for most people
为什么选它 / 目的Why you would pick it / its purpose如果你的瓶颈是显存：直接下 Q6_K（31 GB，保守档）。 如果你的瓶颈是速度：这个仓里没有更快的档位（GGUF 的速度主要由架构与后端决定，不由量化档决定），换量化者意义不大。 如果你的瓶颈是质量：往上走一档到保守档（Q5_K_M / Q6_K）。If VRAM is your constraint: take Q6_K directly (31 GB, Conservative). If speed is your constraint: this repo has nothing faster — GGUF speed is decided by architecture and backend, not by quantisation tier — so switching quantiser will not help. If quality is your constraint: step up one tier to conservative (Q5_K_M / Q6_K).
已知取舍Known trade-offs· 档位极少，没有升降空间，被显存卡住时无路可退Very few tiers, so there is no room to move up or down when VRAM runs out · 无 imatrix，同档位下质量损失可能比带 imatrix 的版本更明显（我们的口径，需实测验证）No imatrix, so quality loss at the same tier may be more visible than in an imatrix build (our reading, to be verified by measurement)
E量化者档案Quantiser profiles选量化者意味着什么what choosing a quantiser means
量化者决定了这个包的档位覆盖、命名规范、是否带 imatrix、什么时候跟进新模型。这一区块说明「选这个量化者的量化意味着什么」。The quantiser decides this build's tier coverage, naming convention, whether it ships imatrix, and how fast it follows new models. This section spells out what choosing that quantiser means.
Position: A long-standing quantiser with the most complete tier naming and documentation.
Characteristics:
Every tier states what scenario it suits, rather than just dumping files
Naming follows the llama.cpp convention — Q4_K_M is the community's Q4_K_M
The same model often ships both static and imatrix directories
How to choose their builds: read the tier-by-tier README in the repo — it is one of the few places that actually explains *why* one tier loses little quality and another loses more.
Note: repo names often carry a Qwen_ or Meta-Llama_ prefix (Hugging Face repo names must be unique, so quantisers prefix the organisation name to disambiguate) — do not miss the prefix when searching.
Position: Author of the Mirostat sampling algorithm, long-running low-bit experimental quantisations.
Characteristics:
Willing to publish very low-bit tiers others dare not (IQ2_XXS, Q2_K and the like)
Focused on the quantisation method itself — tries different imatrix files and different sources for the importance matrix
File names carry suffixes like i1 or h2, marking which version of the importance matrix was used
Who it suits: experiments along the lines of "fit the largest possible model into the smallest possible VRAM", or comparisons of how different importance matrices affect quality.
Who it does not suit: everyday use. Quality loss at those very low bits is usually already noticeable — unless you genuinely have no alternative, do not make it your primary build.
<p class="muted">This publisher ships no description. What follows is the repo’s own metadata (size, licence, file size, last updated) — no editorialising from us.</p>
<p class="muted">This publisher ships no description. What follows is the repo’s own metadata (size, licence, file size, last updated) — no editorialising from us.</p>
<p class="muted">This publisher ships no description. What follows is the repo’s own metadata (size, licence, file size, last updated) — no editorialising from us.</p>
<p class="muted">This publisher ships no description. What follows is the repo’s own metadata (size, licence, file size, last updated) — no editorialising from us.</p>
按更新时间倒序。更新频繁说明该量化者在跟进新版本；长期不更新则基础模型升版后可能不同步。Sorted by most recent update. Frequent updates mean the quantiser is tracking new releases; a long silence means it may have fallen behind the base model.
本站只负责「该下哪一档」。跑起来是另一件事 —— 下面这几个工具自己的文档里写明了支持本地模型（Ollama / vLLM / llama.cpp / LM Studio），点进去看具体怎么接。This site only answers which tier to download. Running it is a different problem — the tools below document local-model support in their own docs (Ollama / vLLM / llama.cpp / LM Studio). Follow the link to see how.
这些条目的本地模型支持情况由它们自己的官方文档佐证；本站不实测。Local-model support for each entry is evidenced by that project official documentation. This site runs no measurements.
关于数据与许可Data and licensing
本站数据全部来自All data here comes from theHugging Face公开 API，采集时间public API, collected2026-10-30，为快照数据。; this is a snapshot.
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，只提供发布页链接。许可字段原样引自发布页， — it only links to release pages. Licence fields are quoted verbatim from them,本站不解读许可条款、不构成法律意见this site does not interpret licence terms or give legal advice，一切以 and everything is governed by发布页原文the legal page为准。.
模型名称均为各自权利人的商标，此处仅作描述性使用。本站与上述厂商无隶属或合作关系。Model names are trademarks of their respective owners and are used here only descriptively. This site is not affiliated with or endorsed by any of the vendors above.