Muse GlimmerMuse Glimmer

<p class="muted">该发布者未提供说明。以下为该仓库的客观元数据（规格、许可、文件体积、更新时间），不含本站的主观解读。</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>

定位：量化覆盖面最广、下载量最高的第三方量化者。

擅长什么：几乎每个主流基础模型都第一时间跟 GGUF，且档位最全——从极限压缩到近似无损一应俱全。提供静态（static）与动态（dynamic，用 UD- 前缀）两套量化。

• 为什么要认 unsloth 版：
• 覆盖率高 → 你想要的档位通常都有
• 命名规范 → Q4_K_M 就是 Q4_K_M，不会同档不同名
• 带 importance matrix（imatrix）→ 量化质量比纯 RTN 更稳
• 官方教程与 LM Studio / llama.cpp 集成路径最顺

注意：unsloth 的动态量化（UD-Q4_K_XL 等）在同档位下体积略大、速度略慢，换来的是质量通常更好。是否值得取决于你的瓶颈是显存还是速度——这是我们的口径，不是实测结论。

官方发布页：https://huggingface.co/unsloth

Position: The broadest quantisation coverage and the highest downloads among third-party quantisers.

What it is good at: publishes GGUF for almost every mainstream base model first, with the most complete tier range — from extreme compression to near lossless. Offers both static and dynamic (the UD- prefix) quantisation.

• Why choose the unsloth build:
• Broad coverage → the tier you want is usually there
• Naming discipline → Q4_K_M means Q4_K_M, with no different name for the same tier
• Ships an importance matrix (imatrix) → quality is steadier than plain RTN
• The smoothest path to official tutorials and LM Studio / llama.cpp integration

Note: unsloth's dynamic quantisation (UD-Q4_K_XL and similar) is slightly larger and slightly slower at the same tier, in exchange for usually better quality. Whether that trade is worth it depends on whether your bottleneck is VRAM or speed — this is our reading, not a measurement.

Release page: https://huggingface.co/unsloth

定位：LM Studio 桌面客户端的官方社区组织，为其 GUI 供包。

• 特点：
• 档位偏保守务实，以「在 LM Studio 里能顺畅跑起来」为目标筛选
• 文件命名与 LM Studio 的模型库索引对得上，下载后能被客户端直接识别
• 也会带 MTP / 投机采样相关的特殊版本

适合谁：如果你用 LM Studio 而不是自己敲 llama-server 命令行，这个组织的包省心程度最高。

不适合谁：如果你要精细控制加载参数（tensor split、n_gpu_layers、上下文分配），社区包往往不针对你的硬件调优，仍要自己调。

发布页：https://huggingface.co/lmstudio-community

Position: The official community organisation behind the LM Studio desktop client, supplying packages for its GUI.

• Characteristics:
• Tiers lean conservative and pragmatic, selected for "runs smoothly inside LM Studio"
• File naming lines up with LM Studio's model library index, so downloads are recognised by the client directly
• Also ships special builds related to MTP / speculative sampling

Who it suits: if you use LM Studio rather than driving llama-server yourself, packages from this organisation save the most fuss.

Who it does not suit: if you need fine control over load parameters (tensor split, n_gpu_layers, context allocation) — community packages are usually not tuned to your hardware, so you still tune them yourself.

Release page: https://huggingface.co/lmstudio-community

定位：老牌量化者，档位命名规范、说明文档最完整。

• 特点：
• 每档位都写清「这个档适合什么场景」，不是只丢一堆文件
• 命名遵循 llama.cpp 约定，Q4_K_M 就是社区通义的那个 Q4_K_M
• 同一个模型常常同时提供静态量化与 imatrix 量化两个目录

怎么选他的包：看仓库里的 README 分档说明，那是少数把「为什么这档质量损失小、那档损失大」写清楚的地方。

注意：他的仓库名常带 Qwen_ Meta-Llama_ 这类前缀（因为 HF 仓库名不允许重复，量化者要加模型组织名消歧），搜仓时别漏掉前缀。

发布页：https://huggingface.co/bartowski

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.

Release page: https://huggingface.co/bartowski

<p class="muted">该发布者未提供说明。以下为该仓库的客观元数据（规格、许可、文件体积、更新时间），不含本站的主观解读。</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">该发布者未提供说明。以下为该仓库的客观元数据（规格、许可、文件体积、更新时间），不含本站的主观解读。</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>

定位：llama.cpp 自己的官方量化发布渠道。

为什么重要：ggml 是 llama.cpp 的底层张量库，ggml-org 发的 GGUF 由项目核心作者把关。遇到新架构（比如新出的 MoE 稀疏模型），ggml-org 通常是第一个支持并发布可用 GGUF 的。

• 特点：
• 档位偏保守，优先保证能正确加载与推理不出错
• 覆盖不一定全（作者精力有限，只跟进自己关心的模型）
• 适合作为「该架构能不能在 llama.cpp 上跑」的权威参考

适合谁：想第一时间在新架构上尝鲜、或者需要确认某个模型是否已被 llama.cpp 支持。

发布页：https://huggingface.co/ggml-org

Position: The official quantisation release channel of llama.cpp itself.

Why it matters: ggml is llama.cpp's underlying tensor library, and ggml-org's GGUFs are vetted by the project's core authors. When a new architecture appears (a new sparse MoE, say), ggml-org is usually the first to support it and publish a usable GGUF.

• Characteristics:
• Tiers lean conservative, prioritising correct loading and inference that does not break
• Coverage is not necessarily complete (limited authorial bandwidth — they track only models they care about)
• Serves as the authoritative reference for "can this architecture run on llama.cpp at all"

Who it suits: people who want to try a new architecture first-hand, or need to confirm whether llama.cpp supports a given model.

Release page: https://huggingface.co/ggml-org

<p class="muted">该发布者未提供说明。以下为该仓库的客观元数据（规格、许可、文件体积、更新时间），不含本站的主观解读。</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>

定位：Mirostat 采样算法作者，长期做低比特实验性量化。

• 特点：
• 愿意发别人不敢发的极低比特档位（IQ2_XXS、Q2_K 这类）
• 关注点在「量化方法本身」——会试不同的 imatrix、不同的重要性矩阵来源
• 文件命名带 i1 h2 之类的后缀，标明用的是哪一版重要性矩阵

适合谁：在做「极小显存里塞下最大模型」这类实验，或者想对比不同重要性矩阵对质量的影响。

不适合谁：日常使用。极低比特档位的质量损失通常已经很明显了，除非你确实没有别的选择，否则不建议作为主力。

发布页：https://huggingface.co/mradermacher

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.

Release page: https://huggingface.co/mradermacher

<p class="muted">该发布者未提供说明。以下为该仓库的客观元数据（规格、许可、文件体积、更新时间），不含本站的主观解读。</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>