MiniMax M3.1 Flash

MiniMax M3.1 Flash

Quick answer: MiniMax M3.1 Flash is MiniMax's October 2026 refresh of M3 — a native multimodal mixture-of-experts model (~428B total, 23B active) for coding, agentic workflows, computer use, and cybersecurity, with a 1M-token context window. It scores 76.2% on BankerToolBench, 73.7% on DeepSWE v1.1, and 61.7% on OSWorld 2.0.

At a Glance

Where MiniMax M3.1 Flash leads

• Agentic coding: 73.7% DeepSWE v1.1, 61.2% NL2Repo-Bench
• Cybersecurity: 82.2% CyberGym
• 1M-token context with sparse attention + DSpark speculative decoding for efficiency

Where it lags

• Terminal-Bench 4.0: 38.1% — below several frontier peers on raw terminal-use tasks
• MiniMax Community License: not permissively open source
• Narrower public benchmark disclosure than some rivals (several vendor-internal evals not independently reproducible)

Best for: Agentic coding and computer-use workloads needing long context and efficient inference over M3.

What MiniMax M3.1 Flash Is

MiniMax M3.1 Flash builds on M3's mixture-of-experts architecture (~428B total parameters, ~23B active per token) with three changes: sparse attention replacing the first three full-attention layers, quantization-aware training of routed experts and attention paths, and DSpark speculative decoding replacing the earlier EAGLE-like draft approach. It accepts text, image, and video input and produces text output, with a context window of up to 1,000,000 tokens.

The model targets agentic use cases — coding, computer-use (OSWorld), tool-calling, and cybersecurity research — rather than pure knowledge/reasoning benchmarks, continuing MiniMax's enterprise-agentic positioning from M3.

Specifications

Pricing

Pricing not yet published by MiniMax for M3.1 Flash at time of writing; available via the MiniMax API and MiniMax Agent.

Context Window

MiniMax M3.1 Flash has a 1,000,000-token context window — enough for large multi-file codebases or long agent trajectories in a single request. Sparse attention and DSpark speculative decoding are designed to keep this long-context setting efficient at inference time.

Public Benchmark Scores

Scores above are reported by MiniMax and shown for context. They are not Benchgen measurements. Figures depend on harness, tools, and effort settings — see the source for methodology.

Excluded vendor scores (not mapped to any local benchmark entity): Automation-Bench v1.0.6 (62.8%) — distinct from the locally-tracked Zapier AutomationBench despite the similar name, no confirmed shared methodology; GDP.pdf AA (20.8) — ambiguous metric scale/label in the vendor table, not confidently mappable to GDPVal-AA v2's Elo scale; LOCA-Bench 256K (60.3%) — no matching local entity; SWE Refactor Bench (23.0%) — distinct from the locally-tracked Scale AI SWE Atlas – Refactoring benchmark, no confirmed shared methodology. Each excluded per the scope-assumption rule rather than guessed.

MiniMax M3.1 Flash vs Alternatives

MiniMax M3.1 Flash trades some raw terminal-use performance (Terminal-Bench 4.0) for stronger cybersecurity and coding-agent results versus its DeepSeek Flash-class rivals.

How MiniMax M3.1 Flash Performs on Real Agent Tasks

The public numbers span enterprise tool-use (BankerToolBench), computer-use (OSWorld 2.0), and long-horizon coding (DeepSWE, NL2Repo-Bench) — a broader agentic footprint than M3's narrower CLAW Bench-only disclosure. A leaderboard score establishes general capability on a fixed task distribution; builders evaluating M3.1 Flash for a specific domain (e.g. banking workflows or vulnerability research) should still validate against their own task set, since harness and tool-access settings materially affect reported scores.

Use MiniMax M3.1 Flash via API

from openai import OpenAI  # MiniMax exposes an OpenAI-compatible API

client = OpenAI(base_url="https://api.minimax.io/v1", api_key="YOUR_API_KEY")

response = client.chat.completions.create(
    model="MiniMax-M3.1-Flash",
    messages=[{"role": "user", "content": "Summarize this contract clause..."}],
)
print(response.choices[0].message.content)

Related Models & Benchmarks

• Compare: MiniMax M3 · MiniMax M2.5
• Benchmarks: DeepSWE · OSWorld 2.0 · CyberGym

Frequently Asked Questions

Specs and scores sourced from MiniMax's official M3.1 Flash model card (October 2026). Pricing cited to the provider where available. Last updated 2026-10-24.