Skip to page content
SAVRN
Search Plan a project

Open-weight model · Text generation

Qwen3.8-27B-Flash-MTP-GGUF

by SuperLogic SuperLogic/Qwen3.8-27B-Flash-MTP-GGUF

Qwen3.8-27B-Flash-MTP-GGUF is an open-weight model for text generation from SuperLogic, released under Apache License 2.0. Its published files total 6.8 GB.

让普通用户进入 0 Token 时代。 不需要租 API、不需要云、不需要买卡。一台 8 GB 显存的普通笔记本,在自己家里跑 27B 级推理模型: 作为对照:同一台机器上,LM Studio 跑 Qwen3.8-27B 的 4-bit 量化,大约 1–2 t/s。 最后一条是关键:MTP 草稿头的 output.weight 是全词表 [5120, 248320](265 MB),每个起草步都要整读一遍。 目标模型验证时仍然用全词表,所以输出分布一个 bit 都不变 —— 只省字节,不伤质量(实测接受率持平)。…

Parameters—
Context—
Weights6.8 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—

Model Card

By SuperLogic, published under apache-2.0, revision 0d0406bac341.

让普通用户进入 0 Token 时代。 不需要租 API、不需要云、不需要买卡。一台 8 GB 显存的普通笔记本,在自己家里跑 27B 级推理模型: 作为对照:同一台机器上,LM Studio 跑 Qwen3.8-27B 的 4-bit 量化,大约 1–2 t/s。 最后一条是关键:MTP 草稿头的 output.weight 是全词表 [5120, 248320](265 MB),每个起草步都要整读一遍。 目标模型验证时仍然用全词表,所以输出分布一个 bit 都不变 —— 只省字节,不伤质量(实测接受率持平)。 上图:开 low 思考档,模型生成的中国象棋棋盘布局。 脚本和二进制在 GitHub —— 刻意分成两个仓,因为脚本得能被人读、被人改: features.ini 不在引擎源码仓里,也不是编译进二进制的。 引擎没有它的内置默认路径, 它是启动时用 --features-file features.ini 读进来的。少了它,自己编出来的二进制跑的是 上游行为,不是我们的调校(比如 spec-draft-vocab-map 那条 24k 起草短名单、kv-vram-cells 的 KV 分箱、spec-draft-sampling = probabilistic)。 复现要点(细节见 runtime 仓的 README): - 采样 --temp 0.6 --top-k 20 --top-p 0.95 - KV 缓存用 q40,长上下文靠 tiered KV(kv-vram-cells)把超出的部分放 pinned host 内存 - 自投机必须带…

Read SuperLogic's full model card

Qwen3.8-27B-Flash-MTP — 27B 三值 + MTP:8 GB 笔记本显卡跑到 70 t/s

让普通用户进入 0 Token 时代。

不需要租 API、不需要云、不需要买卡。一台 8 GB 显存的普通笔记本,在自己家里跑 27B 级推理模型:

模式 实测速度
关思考 70 t/s
编辑模式 100–200 t/s

作为对照:同一台机器上,LM Studio 跑 Qwen3.8-27B 的 4-bit 量化,大约 1–2 t/s。

上图:8 GB 普通笔记本上的实测输出速度。


为什么能做到

27B 从 ~54 GB(FP16)压到 5.79 GB,靠的是四件事叠加:

技术 效果 来源
三值量化 + 配套 CUDA 内核 27B 压到 5.79 GB,单卡可跑 PrismML Bonsai-2 主干 + 本 fork 的内核
CUDA Graph 提速约 1.5× llama.cpp / ggml
MTP 推测解码,深度 3 一次前向多出 3 个 token 本 fork(草稿头内嵌在 gguf 里,blk.64.nextn)
起草词表 248k → 24k 起草一步只读 25.6 MB,省 90% 字节 我们自己的优化

最后一条是关键:MTP 草稿头的 output.weight 是全词表 [5120, 248320](265 MB),每个起草步都要整读一遍。 把起草打分限制在自己语料真正用到的 24,000 行上,每步只读 ~26 MB。 目标模型验证时仍然用全词表,所以输出分布一个 bit 都不变 —— 只省字节,不伤质量(实测接受率持平)。


三值量化没有伤推理能力

三值化压缩的是记忆的精度,不是推理能力 —— 后者可以通过注入更丰富的知识库记忆和图像记忆恢复。

上图:开 low 思考档,模型生成的中国象棋棋盘布局。


完整复现方案

跑起来要三样东西:脚本和配置、引擎二进制、权重。权重就是这个仓库。 脚本和二进制在 GitHub —— 刻意分成两个仓,因为脚本得能被人读、被人改:

拿什么 在哪
启动脚本 llmserver.bat / .sh、features.ini、命令行客户端 github.com/zgsndzq/qwen38-27b-flash-mtp
预编译 Windows 运行时(含 CUDA 运行时 dll,不用装 CUDA Toolkit) Releases
引擎源码 + 构建脚本(想自己编才要) github.com/zgsndzq/llama.cpp-qwen38-flash

features.ini 不在引擎源码仓里,也不是编译进二进制的。 引擎没有它的内置默认路径, 它是启动时用 --features-file features.ini 读进来的。少了它,自己编出来的二进制跑的是 上游行为,不是我们的调校(比如 spec-draft-vocab-map 那条 24k 起草短名单、kv-vram-cells 的 KV 分箱、spec-draft-sampling = probabilistic)。

最快的一条路:

git clone https://github.com/zgsndzq/qwen38-27b-flash-mtp
cd qwen38-27b-flash-mtp
:: 把 Release 的二进制包解压到当前目录,把本仓库两个 gguf 放进 weights\
llmserver.bat

复现要点(细节见 runtime 仓的 README):

  • 采样 --temp 0.6 --top-k 20 --top-p 0.95
  • KV 缓存用 q4_0,长上下文靠 tiered KV(kv-vram-cells)把超出的部分放 pinned host 内存
  • 自投机必须带 --spec-type draft-mtp-adaptive,否则 MTP 草稿头不参与起草,速度掉一半以上
  • 起草短名单 --spec-draft-vocab-map maps/shortlist_24k_mix.vocab.txt

文件

文件 大小 说明
Qwen3.8-27B-Flash-MTP.gguf 5.79 GB 主模型。三值主干 + 内嵌 MTP 草稿头(866 张量),嵌入/输出层用 q4_K_M 保住精度
mmproj-Qwen3.8-27B-Q8_0.gguf 0.59 GB 视觉投影器,要读图才需要

sha256 见本仓库 SHA256SUMS。

只有这两个。 我们本地开发期间试过的其他变体(不同 MTP donor、不同嵌入精度、以及一个没有 MTP 头的版本)都不发布 —— 免得下载者拿到一个"看起来一样但用不了自投机"的文件。


怎么跑

llama-server.exe ^
  -m Qwen3.8-27B-Flash-MTP.gguf ^
  --host 127.0.0.1 --port 18080 --parallel 1 ^
  -c 65536 -ngl 99 -fa on -ctk q4_0 -ctv q4_0 ^
  --spec-type ngram-mod,draft-mtp-adaptive --spec-draft-n-max 4
  • 必须用本 fork 的引擎。PTQ1_0 张量类型和 MTP 草稿上下文都是这个 fork 才有的, 官方 llama.cpp 加载会直接报 unknown tensor type。
  • 视觉:加 --mmproj mmproj-Qwen3.8-27B-Q8_0.gguf --image-min-tokens 1024。
  • 8 GB 卡建议 -c 65536;6 GB 卡降到 32768。

出处与许可

License: Apache 2.0

本仓库是衍生作品,不是原创模型。完整链条(逐层均为 Apache 2.0):

层 出处 做了什么
1 Qwen/Qwen3.8-27B Copyright 2026 Alibaba Cloud 原始模型(27B 混合注意力,约 75% 线性 + 25% 全注意力)
2 prism-ml/Ternary-Bonsai-2-27B-gguf PrismML 矩阵权重转三值(PTQ1_0,ternary g128 + 块内 Hadamard 旋转 + FP16 组缩放)
3 ukisai/Swift-Bonsai-2-GGUF UkisAI reasoning-efficiency 微调(官方报告思考 token 减约 39.8%,基准分持平)
4 MTP 草稿头,来自 ProCreations 本仓库 gguf 的 graft.donor.name 指向它
5 本仓库 + zgsndzq/llama.cpp-qwen38-flash 合并第 3、4 层为一个 GGUF,并按精度重新打包。未重新训练;速度优化与起草短名单是本项目做的

引擎侧(zgsndzq/llama.cpp-qwen38-flash,fork of PrismML-Eng/llama.cpp)为 MIT。

必填署名: "Created using Bonsai by Prism ML."

使用本仓库需同时遵守上述各层的许可条款。完整声明见 NOTICE 与 LICENSE。


引用

@misc{swift-bonsai-2,
  title  = {Qwen3.8-27B-Flash-MTP: 27B ternary (PTQ1\_0) + MTP, 70 tok/s on 8 GB},
  note   = {Derived from Qwen3.8-27B via PrismML Ternary Bonsai 2 and UkisAI Swift Bonsai 2},
  year   = {2026}
}

Identity and Version

Repository
SuperLogic/Qwen3.8-27B-Flash-MTP-GGUF
Publisher
SuperLogic
Task
Text generation
Modality
Text
Library
llama.cpp
Parameters
Not stated by the source
Languages
zh, en
Revision
0d0406bac341fcc23d3d60320112c2db359dee65
First published
2026-10-15
Last updated
2026-10-16

Files and Weights

5 files, 6.8 GB in total. The weights are 2 files totalling 6.8 GB in gguf.

Weights2 files · 6.8 GB
Documentation1 file · 6.9 KB
Other1 file · 245 B
Repository1 file · 1.7 KB
Every file
FileTypeSizeSHA-256
Swift-Bonsai-2-PTQ1_0-mtp-pro-q4km.ggufWeights6.2 GB 311698b7c61b
mmproj-Qwen3.8-27B-Q8_0.ggufWeights629.2 MB 14ce2bf2eca1
README.mdDocumentation6.9 KB —
SHA256SUMSOther245 B —
.gitattributesRepository1.7 KB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
6.8 GB
Download from SuperLogic

Released by SuperLogic through its official repository on Hugging Face. Read the license.

Built From

Memory Requirements

PrecisionWeights in memory
As published6.8 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About Qwen3.8-27B-Flash-MTP-GGUF

Can I use Qwen3.8-27B-Flash-MTP-GGUF commercially?

Yes. Qwen3.8-27B-Flash-MTP-GGUF is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Model · Text generation

opt-125m

AI at Meta

OPT was first introduced in Open Pre-trained Transformer Language Models and first released in metaseq's repository on May 3rd 2022 by Meta AI. Disclaimer: The team releasing OPT wrote an official model card, which is available in Appendix D of the paper. Content from this model card has been written by the Hugging Face team. To quote the first two paragraphs of the official paper OPT was predominantly pretrained with English text, but a small amount of non-English data is still present within the training corpus via CommonCrawl. The model was pretrained using a causal language modeling (CLM) objective. OPT belongs to the same family of decoder-only models like GPT-3. As such, it was…

Open weights other 2,048 tokens transformers

Model · Text generation

Ornith-1.5-9B-GGUF

Ornith

Chirp Chirp! We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement. Ornith-1.5 extends Ornith-1.0 (which was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. For…

Open weights mit transformers

Model · Text generation

Ternary-Bonsai-2-27B-gguf

Prism ML

Full 27B-class reasoning in ternary transformer weights, for llama.cpp (CUDA, Metal, CPU) - \~5.9 GB language model (down from \~54 GB FP16) — full 27B-class reasoning on a standard laptop or a single GPU - 98.2% of FP16 intelligence retained: 84.78 average across 14 thinking-mode benchmarks — far above the conventional IQ2XXS build (72.59) at about 82% of its footprint, and within 0.4 points of UD-Q4KXL at three times the footprint - Retains thinking, reasoning, and agentic behavior deep in the sub-4-bit regime, where conventional low-bit representations collapse: math within half a point of full precision (96.57), coding level with the baseline (89.42), agentic tool calling at 74.92…

Open weights apache-2.0 llama.cpp

Model · Text generation

Ornith-1.5-35B-A3B-GGUF

Ornith

Chirp Chirp! We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement. Ornith-1.5 extends Ornith-1.0 (which was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. For…

Open weights mit transformers

Uncensored Qwen3.8-27B, published as GGUF quantizations with the multi token prediction (MTP) head retained and verified. Refusal behaviour has been substantially reduced, not eliminated. See Measured behaviour for the numbers. Capabilities, training data, and architecture are otherwise unchanged. - Refusal directions removed with Heretic, which co minimizes refusal count against KL divergence from the base model. No handwritten refusal removal code, no finetuning, no additional training data. - Abliteration runs at bf16 (no 4 bit quantization). the resulting LoRA is merged into the bf16 base, so the published weights are not a quantized round trip. - mtp. tensors are copied verbatim from…

Open weights apache-2.0 llama.cpp

Model · Text generation

Ornith-1.0-9B-GGUF

Ornith

Aloha! Today, we are releasing Ornith-1.0, a self-improving family of open-source models for agentic coding. This model card documents Ornith-1.0-9B, the most lightweight member of the Ornith family, designed for efficient single-GPU deployment. Ornith-1.0-9B is a dense ~9B model (≈19 GB in bf16), so it serves comfortably on a single 80GB GPU. The recipes below stand up an OpenAI-compatible server; add --tensor-parallel-size / --tp if you want to shard across more GPUs. For a quick local test (or to script offline generation), load the model directly with Transformers. Make sure you have a recent release installed — see the Transformers installation guide; Ornith-1.0-9B requires…

Open weights mit transformers