LLM AtlasLLM AtlasSearch models

Xiaomi

MiMo-V2-Flash

Open weightopen-weight

Xiaomi's MiMo-V2-Flash: 309B total/15B active MoE with hybrid sliding window attention, Multi-Token Prediction, and 256K context. Scores 94.1 on AIME 2025, 73.4 on SWE-Bench. Trained on 27T tokens with 6x KV-cache reduction.

Last verified: 2026-03-29Confidence: HighSources: 4
textreasoningcodeopen-source
Input price
Unpublished
Output price
Unpublished
Context window
262,144
Max output
32,768
Release date
2025-12-16
Access
open-weight, self-hosted, hosted, api
License
MIT
Last verified
2026-03-29

Capability profile

Radar view of the model's practical strengths. This chart is backed by textual summaries below for crawlability.

Benchmark summary

Scores 94.1 on AIME 2025, 83.7 on GPQA-Diamond, and 73.4 on SWE-Bench Verified with 3x inference speed via MTP.

Strengths

  • • 3x output speed via MTP
  • • 256K context
  • • SOTA reasoning (94.1 AIME)
  • • Strong agentic capability (73.4 SWE-Bench)
  • • MIT license

Trade-offs

  • • 309B total params require significant hardware
  • • Lower than Gemini-3.0 Pro on most benchmarks

Crawlable benchmark analysis

MiMo-V2-Flash is positioned as an open-weight reasoning model with published scores that emphasize its practical fit for buyers evaluating the entry.

Published scores highlight reasoning 86/100, coding 82/100, enterprise readiness 62/100, vision 40/100, speed 82/100, and safety 68/100.

Pricing is not fully published for this entry. With a context window of 262,144 tokens, it supports large-document analysis and retrieval workflows.

Across the tracked benchmark set, MiMo-V2-Flash shows especially strong performance in ARC-Challenge, making it a viable option for teams prioritizing reasoning-heavy use cases.

Sources

Provider and distribution links used to verify this model record.

Last verified: 2026-03-29

Related models

OpenAI

GPT-5.4

OpenAI

OpenAI's GPT-5.4, the most capable and efficient frontier model for professional work. First general-purpose model with native computer-use capabilities. Combines industry-leading coding from GPT-5.3-Codex with improved agentic workflows.

Score 933 sources
textreasoningtool-usevisionapihosted
Context
1,000,000
Input
$0.005/1K tok
Output
$0.02/1K tok
Coverage
Full profile
View analysis

Anthropic

Claude Sonnet 4.6

Claude 4.6

Anthropic's current Sonnet tier for fast frontier reasoning, coding, and long-context agent work.

Score 923 sources
textvisionreasoningcodetool-useapihosted
Context
1,000,000
Input
$0.003/1K tok
Output
$0.02/1K tok
Coverage
Full profile
View analysis

Anthropic

Claude Opus 4.6

Claude 1M

Anthropic's most intelligent Claude model for complex agents, coding, and deep reasoning, with 1M token context and 128K output.

Score 913 sources
textvisionreasoningapihosted
Context
1,000,000
Input
$0.005/1K tok
Output
$0.03/1K tok
Coverage
Full profile
View analysis