Research Vault

MiniMax M3.1-Flash

Properties1
tagshuggingface, text-generation, coding, agentic, multimodal, moe, instruct, tool-use, computer-use, long-context, cybersecurity, china-lab, qat

4 min read

MiniMax M3.1-Flash

One-line: MiniMax’s open-weight GA of M3.1-Flash — a 428B-param / 23B-active native-multimodal MoE built for coding, agentic workflows, computer use and cybersecurity, 1M context, MXFP8 QAT checkpoint with a DSpark draft head; lab-reported DeepSWE v1.1 73.7, Automation-Bench 62.8, CyberGym 82.2, OSWorld 2.0 61.7. ⚠ large — ≈430 GB at MXFP8 across 48 shards (server-class only). ⚠ no SWE-bench Verified / LiveCodeBench / Aider / BFCL published yet; MiniMax’s card uses DeepSWE v1.1 + Terminal-Bench 4.0 as its coding proxies.

Weights landed 2026-10-23 03:37Z in a single Init MiniMaxAI/MiniMax-M3.1-Flash commit (48 safetensors shards + dspark/ draft head); the repo itself was created 2026-10-17 but was private until the drop (the 2026-10-22 run’s ?search=MiniMax-M3.1 returned 0). The API-only M3.1-Flash launched 2026-09-27 as a preview; this is the open-weights release. Verified license: MINIMAX COMMUNITY LICENSE (file in repo) — see caveat below.

⚠ All benchmark numbers below are lab-reported from the model card’s comparison table (evaluated with third-party harnesses: mini-SWE-agent 2.4.5, Claude Code 2.1.x, Codex 0.149.1, OpenCode 1.17.18 — stated per row in the card’s Evaluation Notes). BenchLM’s independent compilation ranks the model Coding #17/152 and Agentic #19/131, but its rows are also flagged “provider exact” — i.e. still provider-sourced. No independent re-run found as of 2026-10-24.

Coding benchmarks

BenchmarkScoreSource
HumanEval—not published
HumanEval+—not published
MBPP—not published
LiveCodeBench—not published
Aider polyglot—not published
SWE-bench Verified—not published
SWE-bench Multilingual—not published
DeepSWE v1.173.7lab — mini-SWE-agent 2.4.5, 512K ctx; vs DeepSeek V4.1 Flash-1017 77.0, V4.1 Flash 74.2, Opus 5 74.0
Terminal-Bench 4.038.1lab — Codex 0.149.1; vs DeepSeek V4.1 Flash-1017 42.3, Opus 5 51.8
NL2Repo-Bench61.2lab — Claude Code 2.1.233, 512K ctx, reward-hacking 0; vs DeepSeek V4.1 Flash-1017 68.2
SWE Refactor Bench23.0lab — Claude Code 2.1.220, 200K ctx; best in row (vs Opus 5 31.0 — trailing)
PostTrain Bench v1.139.3lab — Claude Code 2.1.251, 512K ctx
SciCode—not published
MLE-bench Lite—not published

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL—not published
τ-Bench—not published
ToolACE—not published
GAIA—not published
Automation-Bench v1.0.662.8lab — 600 public tasks; best in row (vs DeepSeek V4.1 Flash-1017 60.5, Opus 5 50.3)
BankerToolBench76.2lab — OpenCode 1.17.18, MiniMax-M2.7 judge; ties Opus 5 76.2
OSWorld 2.061.7lab — 0808 full release; vs Opus 5 70.6, GPT-5.6 Sol 65.7
CyberGym82.2lab — Claude Code 2.1.263, 1M ctx; vs GLM-5.3 84.5
Agents’ Last Exam29.6lab — official Overall setting, full-set leaderboard rows only (blanks = no official result, not zero)
LOCA-Bench 256K60.3lab — official implementation; best in row (vs GPT-5.6 Sol 60.0)
BabyVision (with tools)91.0lab — Claude Code 2.1.285; vs DeepSeek V4.1 Flash-1017 92.5
CharXiv RQ82.5lab — GPT-5.5 judge

Reasoning (context only)

BenchmarkScoreSource
MMLU / MMLU-Pro—not published
GPQA Diamond—not published
MATH—not published
AIME 2025 / 2026—not published
HLE (without tools)43.4lab — text-only subset (2,158 questions)

Efficiency

  • Active params: 23B active / 428B total (MoE). Sparse attention replaces the first three full-attention layers; routed experts and attention paths are quantized during training (QAT).
  • Context length: 1,000,000 (1M) — usable length and throughput depend on GPU memory and serving config.
  • VRAM: ⚠ ≈430 GB at MXFP8 (48 shards; the shipped checkpoint is QAT MXFP8, not BF16 — MiniMaxAI has no BF16 sibling repo, so treat MXFP8 as the canonical checkpoint rather than a quant re-release). Server-class only; no single-80GB path.
  • Runtime: SGLang (cookbook), vLLM (recipe), TokenSpeed recipes at launch; Transformers / KTransformers / unsloth / ATOM “in progress”. Multi-chip via FlagOS. Custom code required (model_type: minimax_m3_vl, MiniMaxM3SparseForConditionalGeneration, ships configuration_minimax_m3_vl.py + processing_minimax.py). Chat template controls reasoning_effort (low→max); recommended temp 1.0 / top_p 0.95. DSpark draft head included for speculative decoding. Text + image + video in, text out.
  • License caveat: minimax-community — free for non-commercial use; commercial use requires prominently displaying “Built with MiniMax M3.1-Flash” plus a one-time notice to api@minimax.io, and prior written authorization above US$20M/year revenue. Prohibited-uses appendix (military, minors, disinformation, discrimination). Not a bare Apache/MIT drop — eyes open for anything shipping to customers.

What makes it notable

MiniMax’s first open-weights release at this scale tier — 428B/23B active, natively multimodal, 1M context — and it lands squarely in the agentic-coding lane rather than chat: the card’s headline rows are DeepSWE v1.1 73.7, Automation-Bench 62.8 (best in its comparison row), BankerToolBench 76.2, OSWorld 2.0 61.7 for computer use, and CyberGym 82.2. Against its nearest open peer, DeepSeek-V4.1-Flash-1017 (already tracked), it wins Automation-Bench (62.8 vs 60.5) and LOCA-Bench 256K (60.3 vs —) but trails on DeepSWE v1.1 (73.7 vs 77.0), Terminal-Bench 4.0 (38.1 vs 42.3) and NL2Repo (61.2 vs 68.2) — so it is a strong-but-not-leading coding model whose edge is breadth (vision/video input, 1M ctx, computer-use, cyber). BenchLM (independent compilation, provider-sourced rows) places it Coding #17/152 and Agentic #19/131, capability 67.1 (#27/227). Two real caveats for a local runner: the shipped checkpoint is MXFP8 QAT at ≈430 GB (frontier memory regardless of the 23B active compute), and the community license adds attribution + a revenue threshold where DeepSeek’s MIT does not.

See also

Graph View