跳到主内容
@wquguru
精选88MarkTechPost(RSS)模型发布/更新多源精选 ×2

JetBrains发布Mellum2.1:面向编程Agent的12B

JetBrains Releases Mellum2.1: A 12B MoE Open Model for Coding Agents

原文
发到 X
推荐理由

小参数MoE在编程Agent场景实现显著突破,RL在真实仓库的训练路径对自建Agent有直接参考价值,建议关注其量化部署方案。

JetBrains has released Mellum2.1, an open model built for coding agents and fast sub-agents. Mellum2.1 is a 12B mixture-of-experts thinking model from JetBrains that activates 2.5B parameters per token. It ships under Apache 2.0 on Hugging Face. The architecture is unchanged from Mellum2. The upgrade comes almost entirely from reinforcement learning (RL) in real software environments. The result is a small, self-hostable model that explores a repository, edits files, and checks its own changes.

JetBrains 发布了 Mellum2.1,这是一款专为编码代理和快速子代理构建的开源模型。Mellum2.1 是 JetBrains 推出的 120 亿参数混合专家(MoE)推理模型,每个 token 激活 25 亿参数。该模型以 Apache 2.0 许可证在 Hugging Face 上发布。其架构与 Mellum2 保持一致。此次升级几乎完全来自真实软件环境中的强化学习(RL)。结果是一个小型、可自托管的模型,能够探索代码仓库、编辑文件并检查自身的更改。

TL;DR

TL;DR(太长不看)

  • Size: 12B total, 2.5B active (64 experts, 8 active), 131,072-token context
  • Runs on: your own GPUs via vLLM or SGLang (speed tested on 1 NVIDIA H200). GGUF builds start at 7.0 GB for llama.cpp, Ollama and LM Studio.
  • Performance: beats Mellum2 on 15 of 17 listed benchmarks; wins 5 of 17 against Qwen3.5-9B
  • Best: 82.0 on LiveCodeBench v6, ahead of Qwen3.5-9B (75.4) and Gemma 4 E4B (69.4)
  • Worst: 17.4 on Terminal-Bench 2.1, below Qwen3.5-9B (21.7)
  • Bottom line (best): strong coding scores and high throughput from only 2.5B active parameters.
  • Bottom line (worst): still trails Qwen3.5-9B on hard agentic software tasks and knowledge.
  • 规模:总计 120 亿参数,25 亿活跃参数(64 个专家,每次激活 8 个),上下文长度 131,072 个 token
  • 运行环境:通过 vLLM 或 SGLang 在你的 GPU 上运行(速度已在 1 块 NVIDIA H200 上测试)。GGUF 格式构建版本从 llama.cpp、Ollama 和 LM Studio 的 7.0 GB 起。
  • 性能:在列出的 17 项基准测试中有 15 项优于 Mellum2;在与 Qwen3.5-9B 的对比中赢得 5 项。
  • 最佳表现:LiveCodeBench v6 得分 82.0,领先于 Qwen3.5-9B(75.4)和 Gemma 4 E4B(69.4)
  • 最差表现:Terminal-Bench 2.1 得分 17.4,低于 Qwen3.5-9B(21.7)
  • 总结(最佳):凭借仅 25 亿活跃参数,拥有强大的编码分数和高吞吐量。
  • 总结(最差):在复杂的代理式软件任务和知识方面仍落后于 Qwen3.5-9B。

What is Mellum2.1?

什么是 Mellum2.1?

Mellum2.1 is the next version of Mellum2 Thinking, which JetBrains open-sourced in June 2026. The released checkpoint is Mellum2.1-12B-A2.5B-Thinking. It is a reasoning model that emits its chain of thought before answering. JetBrains targets 3 uses: agent worker, general reasoning assistant, and private self-hosted deployment.

Mellum2.1 是 Mellum2 Thinking 的下一个版本,JetBrains 于 2026 年 6 月将其开源。发布的检查点为 Mellum2.1-12B-A2.5B-Thinking。这是一个推理模型,会在回答之前输出其思维链。JetBrains 瞄准三个用途:代理工作节点、通用推理助手以及私有自托管部署。

How does the Mellum2.1 architecture work?

Mellum2.1 的架构是如何工作的?

The model has 28 layers and 64 experts. A router activates 8 experts per token. Attention uses grouped-query attention with 32 query heads and 4 KV heads. 3 of every 4 layers use a 1,024-token sliding window. Context length is 131,072 tokens, and the vocabulary has 98,304 tokens. Weights ship in bfloat16.

该模型有 28 层和 64 个专家。路由器为每个 token 激活 8 个专家。注意力机制使用分组查询注意力(GQA),包含 32 个查询头和 4 个 KV 头。每 4 层中有 3 层使用 1,024 个 token 的滑动窗口。上下文长度为 131,072 个 token,词表大小为 98,304 个 token。权重以 bfloat16 格式发布。

How was Mellum2.1 trained?

Mellum2.1 是如何训练的?

Almost all of the new work went into post-training. RL moved from a short final stage to the main part of training. JetBrains added RL tasks in math, competitive programming, science, tool use, and software engineering. It filtered open RL datasets for broken tests, unverifiable answers, and tasks that were too easy or impossible. For software engineering, the model trains in real repositories with a shell and file-editing tools. It is rewarded when the tests pass. Training launched millions of sandboxes across thousands of environments.

几乎所有的新工作都投入到了后训练阶段。强化学习(RL)从短暂的最终阶段转移到了训练的主要部分。JetBrains 在数学、竞技编程、科学、工具使用和软件工程领域增加了 RL 任务。它过滤了开放 RL 数据集中存在损坏测试、答案不可验证以及任务过于简单或不可能完成的数据。对于软件工程,模型在带有 shell 和文件编辑工具的真实仓库中进行训练。当测试通过时,它会获得奖励。训练启动了数百万个沙箱,跨越数千个环境。

How does Mellum2.1 perform on benchmarks?

Mellum2.1 在基准测试中的表现如何?

JetBrains evaluated Mellum2.1, Mellum2, Qwen3.5-9B and Gemma 4 E4B with one pipeline in thinking mode. All scores are self-reported by JetBrains.

JetBrains 在思考模式下使用同一套流水线对 Mellum2.1、Mellum2、Qwen3.5-9B 和 Gemma 4 E4B 进行了评估。所有分数均由 JetBrains 自行报告。

The largest jump is agentic coding. SWE-bench Verified rose from 2.0 to 47.0. SWE-bench Pro rose from 0.0 to 28.0. Terminal-Bench 2.1 rose from 0.6 to 17.4. Agentic runs used the open-source Pi v0.73.1 harness with a 114K-token context.

增幅最大的是代理式编码(agentic coding)。SWE-bench Verified 从 2.0 升至 47.0,SWE-bench Pro 从 0.0 升至 28.0,Terminal-Bench 2.1 从 0.6 升至 17.4。代理式运行使用了开源的 Pi v0.73.1 基准测试框架,上下文长度为 114K token。

Mellum2.1 leads the group on LiveCodeBench v6 (82.0), HumanEval+ (91.5), MBPP+ (79.4) and BFCL v4 (62.3). Qwen3.5-9B still leads on SWE-bench Verified (50.0), SWE-bench Pro (38.0), AIME 25/26 (86.7) and GPQA Diamond (77.8). Safety also improved: HarmBench fell from 21.5 to 8.5, where lower is better.

Mellum2.1 在 LiveCodeBench v6(82.0)、HumanEval+(91.5)、MBPP+(79.4)和 BFCL v4(62.3)上领先。Qwen3.5-9B 仍在 SWE-bench Verified(50.0)、SWE-bench Pro(38.0)、AIME 25/26(86.7)和 GPQA Diamond(77.8)上保持领先。安全性也有所提升:HarmBench 从 21.5 降至 8.5,数值越低越好。

Pipelines are important to consider. Qwen’s own card lists 65.6 on LiveCodeBench v6 and 81.7 on GPQA Diamond. JetBrains measured 75.4 and 77.8 for the same model.

流水线是重要的考量因素。Qwen 官方卡片显示其在 LiveCodeBench v6 上得分为 65.6,在 GPQA Diamond 上为 81.7。而 JetBrains 对同一模型的测量结果分别为 75.4 和 77.8。

How fast is Mellum2.1?

Mellum2.1 的速度有多快?

Post-training left the architecture untouched, so speed matches Mellum2. On 1 H200 under heavy load, JetBrains says Mellum2.1 serves almost 2x the tokens of Qwen3.5-9B. For a single request, multi-token prediction (MTP) makes it about 1.6x faster. The MTP head for vLLM speculative decoding is listed as coming soon.

后训练未改动架构,因此速度与 Mellum2 一致。JetBrains 表示,在单张 H200 GPU 高负载下,Mellum2.1 的服务吞吐量约为 Qwen3.5-9B 的两倍。对于单个请求,多令牌预测(MTP)使其速度提升约 1.6 倍。vLLM 推测解码所需的 MTP head 预计即将推出。

How do you run Mellum2.1 locally?

如何在本地运行 Mellum2.1?

The full model serves on vLLM with --reasoning-parser qwen3. Tool calling adds --enable-auto-tool-choice --tool-call-parser hermes. JetBrains recommends temperature 0.6, top_p 0.95 and top_k 20.

通过 vLLM 运行完整模型时需添加参数 --reasoning-parser qwen。启用工具调用需添加 --enable-auto-tool-choice --tool-call-parser hermes。JetBrains 建议设置 temperature 为 0.6,top_p 为 0.95,top_k 为 20。

The JetBrains release notes mentions that GGUF builds are in progress. A GGUF repository already lists 5 files:

JetBrains 发布说明中提到 GGUF 构建正在进行中。一个 GGUF 仓库已列出 5 个文件:

  • BF16: 24.3 GB (reference)
  • Q8_0: 12.9 GB, 96.1% top-token match
  • Q6_K: 10.9 GB, 93.9% top-token match
  • Q4_K_M: 8.1 GB, 88.0% top-token match (recommended)
  • MXFP4_MOE: 7.0 GB, 85.6% top-token match
  • BF16:24.3 GB(参考值)
  • Q8_0:12.9 GB,首 token 匹配率 96.1%
  • Q6_K:10.9 GB,首 token 匹配率 93.9%
  • Q4_K_M:8.1 GB,首 token 匹配率 88.0%(推荐)
  • MXFP4_MOE:7.0 GB,首 token 匹配率 85.6%

Mellum2.1 vs Qwen3.5-9B vs Gemma 4 E4B

Mellum2.1 与 Qwen3.5-9B 及 Gemma 4 E4B 对比

FeatureMellum2.1Mellum2 ThinkingQwen3.5-9BGemma 4 E4B
ArchitectureMoE, 64 experts, 8 activeMoE (same as 2.1)Hybrid Gated DeltaNet + gated attentionDense with per-layer embeddings
Total params12B12B9B (language model)8B with embeddings
Active / effective params2.5B active2.5B active9B (dense)4.5B effective
Context131,072131,072262,144 native, up to 1,010,000128K
ModalityTextTextText, image, videoText, image, audio
LicenseApache 2.0Apache 2.0Apache 2.0Apache 2.0
LiveCodeBench v6*82.069.475.469.4
SWE-bench Verified*47.02.050.023.0
Terminal-Bench 2.1*17.40.621.73.4
BFCL v4*62.349.658.552.5
GPQA Diamond*64.651.077.853.1
AIME 25/26*83.360.186.745.0
特性Mellum2.1Mellum2 思维模型Qwen3.5-9BGemma 4 E4B
架构MoE,64 个专家,8 个激活MoE(与 2.1 相同)混合门控 DeltaNet + 门控注意力稠密结构,含逐层嵌入
总参数量12B12B9B(语言模型)8B(含嵌入)
激活/有效参数量2.5B 激活2.5B 激活9B(稠密)4.5B 有效
上下文长度131,072131,072原生 262,144,最高可达 1,010,000128K
模态文本文本文本、图像、视频文本、图像、音频
许可证Apache 2.0Apache 2.0Apache 2.0Apache 2.0
LiveCodeBench v6*82.069.475.469.4
SWE-bench Verified*47.02.050.023.0
Terminal-Bench 2.1*17.40.621.73.4
BFCL v4*62.349.658.552.5
GPQA Diamond*64.651.077.853.1
AIME 25/26*83.360.186.745.0

*Benchmark rows: JetBrains’ shared pipeline, thinking mode, from the Mellum2.1 model card. Vendors’ own cards report different numbers for Qwen3.5-9B and Gemma 4 E4B.

*基准测试行数据:JetBrains 共享管道,思维模式,源自 Mellum2.1 模型卡片。各厂商自己的模型卡片报告了 Qwen3.5-9B 和 Gemma 4 E4B 的不同数值。

Key Takeaways

关键要点

  • Mellum2.1 is a 12B MoE thinking model with 2.5B active parameters, under Apache 2.0.
  • RL in real repositories lifted SWE-bench Verified from 2.0 to 47.0.
  • It leads the tested group on LiveCodeBench v6 (82.0) and BFCL v4 (62.3).
  • Qwen3.5-9B still wins on SWE-bench Pro, GPQA Diamond and AIME.
  • A 7.0 GB to 8.1 GB GGUF makes local coding sub-agents practical.
  • Mellum2.1 是一个拥有 2.5B 激活参数的 12B MoE 思维模型,采用 Apache 2.0 许可证。
  • 在真实代码仓库中进行的强化学习将 SWE-bench Verified 的得分从 2.0 提升至 47.0。
  • 它在 LiveCodeBench v6(82.0)和 BFCL v4(62.3)上领先于被测试的其他模型。
  • Qwen3.5-9B 仍在 SWE-bench Pro、GPQA Diamond 和 AIME 上保持领先。
  • 7.0 GB 至 8.1 GB 的 GGUF 格式使得本地编码子智能体变得切实可行。

Check out the model weights, GGUF builds, technical blog and the Mellum2 technical report. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

查看模型权重、GGUF 构建版本、技术博客以及 Mellum2 技术报告。所有功劳归于本项目的研究人员。此外,欢迎在 Twitter 上关注我们,别忘了加入我们有超过 15 万成员的 ML SubReddit 并订阅我们的新闻通讯。等等!你在 Telegram 上吗?现在你也可以在 Telegram 上加入我们。

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

关联讨论

同一事件的更多信源

相似阅读

关联信息,但可能不是同一事件