Hacktoberfest 2026:维护者为十月标记出来的 issue,仍然开放、适合新手。 浏览 Hacktoberfest issue

[M3 Request] Demand for a localized "M3-Mini" / Flash variant optimized for 128GB RAM environments

未关闭
#15 1 条评论 6 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
25/100
Issue 类型
功能
描述清晰度
需要澄清
活跃度
冷清

调研方向

该 issue 未指明任何仓库文件、测试或入口点;请从链接的 Local Deployment Guide 开始,并将其中的内存约束与现有的 M3 部署产物进行比较。要完成这项工作,需要就一个更小的模型变体达成一致,并定义内存、延迟和质量目标以及验证方式,因此这属于设计工作,而不是一个可独立完成的首次贡献。

由索引模型根据 Issue 内容生成。

描述

enhancement
Capability area

Deployment / inference

What does M2.7 fail to do for you?

Even with the recent local optimization frameworks, running MiniMax M2.7 on standard 128GB workstation configurations (such as standard Mac Studio setups or dual-GPU rigs totaling 128GB unified/VRAM) is heavily bottlenecked.

The 4-bit MXFP4 and standard 4-bit quantizations hover right around 122GB–129GB. This leaves zero headroom for system memory or the massive KV cache needed for long-context workloads. Running a 4-bit variant often results in active OOM faults or severe context capping (failing to go past low-depth generation), rendering the model's signature deep-context capabilities unusable locally without multiple nodes or cluster setups.

Example Workload: Parsing an entire code repository or a 300k token documentation library locally via vLLM or llama.cpp for multi-step agentic execution.

What would "good" look like in M3?

A native "M3-Mini", "M3-Flash", or a highly-optimized smaller dense variant that occupies an unquantized footprint small enough to comfortably run quantized (e.g., 4-bit or 8-bit) inside a 128GB RAM envelope while preserving a functional KV-cache headroom.

Performance expectations:

Memory Footprint: Quantized model size staying well under 80GB–90GB to dedicate the remaining 38GB+ to the MiniMax Sparse Attention (MSA) context buffer.

Latency: Fast token-per-second decoding speeds (targeting 40-50 tok/s on mid-tier consumer/prosumer clusters).

Quality: Retention of M3’s core agentic capabilities (tool calling, SWE-bench coding depth) even if broad knowledge capacity is scaled down.

References

The official MiniMax Local Deployment Guide explicitly indicates that 4-bit variants require a minimum of 136GB–140GB of memory headroom to function safely, excluding a 128GB machine entirely without aggressive 3-bit performance degradation.

Competitor examples such as Qwen 3.6 35B / 27B combinations or DeepSeek V4 Flash variants, which provide explicit tiers for local deployment footprints.

主要语言
没有语言数据
星标
489
派生
59
PR 合并指标
30 天内没有已合并 PR

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

MiniMax-AI/MiniMax-M3 的其他 Issue

查看 MiniMax-AI/MiniMax-M3 的全部 Issue

相似的 Issue

更多 AI Infra & Agents Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。