xiaomimimo/mimo-v2-flash

1. Repository role

This repo is not a full implementation codebase; it is a model release and technical-report hub for MiMo-V2-Flash. The main artifacts are the README, the PDF report, and figures that document architecture, benchmarks, and deployment guidance.

2. Primary product

The released model family is MiMo-V2-Flash and MiMo-V2-Flash-Base. Both are 309B-parameter MoE language models with 15B active parameters and a 256k context window. The README positions them as high-speed reasoning/agentic models rather than general-purpose application code.

3. Model architecture boundary

The model itself is described as having two major internal subsystems:

- Hybrid attention: interleaved sliding-window attention and global attention for long-context efficiency.

- Multi-Token Prediction (MTP): a lightweight prediction head used to accelerate generation.

These are model-level components, not separate services in this repo, but they define the core inference behavior.

4. Long-context design

The architecture emphasizes long-context throughput by reducing KV-cache pressure through aggressive sliding-window attention, with periodic global attention to preserve quality. The diagram in the README should reflect this as an internal model boundary that trades memory for context scale.

5. Post-training pipeline

The report describes a post-training stack centered on:

- MOPD (Multi-Teacher On-Policy Distillation)

- agentic RL

These are training-time subsystems that likely sit upstream of the released checkpoints. They explain how the base model becomes the instruction-tuned/agentic model.

6. Training and RL infrastructure

The technical report references a sizable reinforcement-learning infrastructure built on SGLang, Megatron-LM, Kubernetes, and Ray. Important architectural pieces include rollout routing replay, request-level prefix caching, fine-grained scheduling, and tool management. In an architecture diagram, these belong to an offline training/rollout plane, separate from end-user inference.

7. Agent/tool execution loop

The README highlights multi-turn tool use, tool-call parsing, and the need to persist reasoning_content across turns. This indicates a conversational loop where the model may emit both natural language and structured tool calls, then continue with tool results as additional context.

8. Inference and serving path

The published serving stack is SGLang. The example launch config shows an inference server with:

- tensor/pipeline/data parallelism

- FP8 mixed precision

- speculative decoding

- MTP enabled

- tool-call and reasoning parsers

The external interface is an OpenAI-style /v1/chat/completions endpoint.

9. Public access and distribution

The repo links to Hugging Face model downloads, a MiMo Studio interactive playground, and a MiMo API Platform. These are the main outward-facing distribution and demo surfaces; the repo itself is the documentation source of truth.

10. Artifacts in this repository

The non-text assets are only supporting visuals:

- architecture/performance figures under figures/

- WeChat contact images under figures/wechat_group/

- the technical report in paper.pdf

There is no source tree for training, serving, or evaluation code here.

11. Evaluation/reporting layer

The README contains extensive benchmark tables for base and post-trained models across reasoning, math, code, Chinese, multilingual, and long-context tasks. This is documentation-only evaluation output, not a runnable benchmarking subsystem inside the repo.

12. Diagramming takeaway

For an architecture diagram, treat this repo as a model release package with three main planes:

- training/post-training pipeline upstream,

- model inference/serving stack in the middle,

- external clients/platforms downstream.

The repository itself anchors the public documentation, model weights distribution links, and deployment contract.