Xiaomi MiMo-V2-Flash — 309B/15B MoE with 5:1 SWA hybrid attention, 73.4 SWE-Bench Verified

TL;DR

Xiaomi announces MiMo-V2-Flash, an open-source MoE with 309B total / 15B active parameters, marketed as “faster than fast” for agentic AI. Architecture is a 5:1 interleaved sliding-window-attention (128-window) + global-attention hybrid with 256K context. The launch claims parity with DeepSeek-V3.2 on general benchmarks at “a fraction of the latency”, new open-source SOTA on SWE-Bench Verified (73.4%) and SWE-Bench Multilingual (71.7%), and 150 output tokens/s decoding throughput with Day-0 SGLang support. No technical report or detailed benchmark table is linked in the tweet itself; the blog and GitHub links land on JS-rendered pages (blog body empty at fetch time) and the stale MiMo-7B repo respectively.

Key claims

• 309B total / 15B active parameter MoE (~4.9% activation ratio) [tweet].
• 5:1 interleaved hybrid attention — five SWA layers (128-token window) per global-attention layer; 256K context window [tweet].
• Matches DeepSeek-V3.2 on general benchmarks at “a fraction of the latency” (no comparison table in tweet) [tweet].
• SWE-Bench Verified 73.4% and SWE-Bench Multilingual 71.7%, claimed new open-source SOTA [tweet].
• 150 output tokens/s decoding speed with Day-0 SGLang serving support [tweet].
• Released on Hugging Face under XiaomiMiMo/MiMo-V2-Flash with a blog post and technical-report links [tweet].

Method

A tweet-as-artifact filing — the only structured detail is the architecture summary (hybrid SWA+global at 5:1 ratio with 128-token windows, 256K context) and the activation ratio (15B / 309B ≈ 4.9% active). The linked technical-report URL points to the github.com/XiaomiMiMo/MiMo repo, which at fetch time still hosts the older MiMo-7B reasoning model README (arxiv:2505.07608) rather than V2-Flash documentation; the blog page (mimo.xiaomi.com/blog/mimo-v2-flash) is JS-rendered and returned an empty body. Numbers and recipe details below are sourced from the tweet text only.

Results

Tweet headline numbers, no peer table:

• SWE-Bench Verified: 73.4%
• SWE-Bench Multilingual: 71.7%
• Output speed: 150 tokens/s (SGLang)
• Context: 256K
• General benchmarks: claimed parity with DeepSeek-V3.2

For external context: the Step 3.5 Flash benchmark table (Step 3.5 Flash) places “MiMo-V2 Flash” at a normalized decoding-cost ratio of 1.2× at 128K context vs Step 3.5 Flash’s 1.0× baseline, putting MiMo-V2-Flash on the same low-active-MoE Pareto frontier as Step 3.5 Flash (11B-active / 196B-total), DeepSeek V3.2 (37B-active), GLM-4.7 (32B-active), Kimi K2 Thinking, and MiniMax M2.1.

Why it’s interesting

MiMo-V2-Flash is another datapoint on the open-MoE-with-low-activation-ratio agentic-SWE Pareto frontier that Agentic Software Engineering is tracking — its 15B/309B ≈ 4.9% activation puts it close to Qwen3-Coder-Next (3B/80B ≈ 3.75%) and Step 3.5 Flash (11B/196B ≈ 5.6%), and its 73.4 SWE-Bench Verified is in the same neighborhood as Step 3.5 Flash’s 74.4. It contrasts with Step 3.5 Flash on the cost-vs-capability axis — Step 3.5 Flash reports MiMo-V2-Flash at only 1.2× its own decoding cost at 128K context, the closest competitor on that table by a wide margin — and complements Qwen3-Coder-Next Technical Report by adding a much larger total-parameter (309B vs 80B) point at a similar activation count. The 5:1 SWA-to-global hybrid attention ratio with 128-token windows is a more aggressive sparsity choice than Step 3.5 Flash’s 3:1 SWA:Full ratio, worth tracking as the hybrid-attention design space for long-context agentic models gets explored. Filed as a tweet rather than as a technical-report page because the linked tech-report URL still resolves to the old MiMo-7B README at fetch time.

See also

• Step 3.5 Flash — Step 3.5 Flash is the explicit peer that tabulates MiMo-V2-Flash’s normalized decoding cost (1.2× vs 1.0×); same low-active-MoE + hybrid-attention design point at a different scale.
• Qwen3-Coder-Next Technical Report — Qwen3-Coder-Next defines the agentic-SWE post-training recipe the open SWE-Bench leaderboard is converging on; MiMo-V2-Flash is the next leaderboard datapoint without a recipe disclosure.
• Introducing GPT-5.3-Codex — Closed-source SWE-Bench Pro upper bound; MiMo-V2-Flash reports Verified+Multilingual rather than Pro, so direct comparison is not yet possible.
• DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models — DeepSeek-V3.2 is the explicit parity claim in the tweet (“matches DeepSeek-V3.2 on general benchmarks at a fraction of the latency”).
• Open foundation-model releases — Another open-MoE coordinated launch; adds the 309B-total-parameter datapoint to the open-foundation-releases scale ladder.
• Agentic Software Engineering — MiMo-V2-Flash sits on the cluster’s SWE-Bench × active-parameters Pareto frontier alongside Qwen3-Coder-Next, GLM-4.7, MiniMax M2.1, Step 3.5 Flash.