On June 1, MiniMax released M3 — their first model built on a completely new attention architecture rather than the Sparse MoE foundation that defined the M2.x line. The headline claims are competitive: 1M token context, 59.0% on SWE-Bench Pro, and inference speeds that make the 1M window usable in production rather than theoretical.
Update — June 14, 2026:
- Open weights shipped. Weights are now available at huggingface.co/MiniMaxAI/MiniMax-M3 and via Ollama (
ollama pull minimax-m3). The MSA technical report was also published (arXiv, June 11). - Promo pricing ended June 8. Standard rate is now $0.60/M input, $2.40/M output for calls up to 512K tokens. Long-context calls (>512K input) are billed at 2x the standard rate — $1.20/M input, $4.80/M output — not the 4x figure an earlier version of this article stated. MiniMax’s pricing structure and the OpenRouter listing both confirm the 2x long-context multiplier — factor it into any cost projection that approaches or exceeds the 512K threshold.
- Updated benchmark: 80.5% SWE-bench Verified. This is MiniMax’s own figure from its June 11 technical report — not an independent third-party evaluation, despite how an earlier version of this article framed it. It’s still a meaningful upgrade from the launch-day claim of 59.0% on SWE-Bench Pro (different benchmark variant, also self-reported). For comparison, Claude Opus 4.8 scores 88.6% on SWE-bench Verified. The gap is real, and both figures are self-reported by their respective labs — independent reproduction of MiniMax’s 80.5% is still pending.
- Baseline caveat confirmed. MiniMax compared M3 against Claude Opus 4.7 in launch materials. By the time M3 shipped June 1, Anthropic had already released Opus 4.8 on May 28. The coding gap vs. current-generation Claude is wider than MiniMax’s stated comparison suggests.
- License: Available on the HuggingFace model card — verify commercial use terms before committing to self-hosted deployment, especially given the M2.7 license tightening precedent. Community discussion on the model card notes the M3 license is more permissive than M2.7’s but that “commercial use” is still ambiguously defined.
This is the full breakdown: what M3 actually is, why the architecture change matters, what the benchmark claims mean and where to be skeptical, how to access it, and who should care.
The Architecture Shift: From Sparse MoE to MSA
The M2.x series — M2, M2.1, M2.5, M2.7 — all shared a core architecture: 229.9 billion total parameters with 9.8 billion active, using Sparse Mixture of Experts (SparseMoE) with 256 experts and 8 activated per token. The approach was solid. M2.5 earned a 4.5/5 in our review for reaching frontier-adjacent benchmark performance at competitive pricing.
M3 is built on something different: MiniMax Sparse Attention (MSA).
MSA is a new attention mechanism designed specifically for long-context inference efficiency. The key claim from MiniMax’s technical announcement:
| Metric | Improvement over prior gen |
|---|---|
| Compute at 1M token context | 1/20th the cost |
| Prefill speed | 9x faster |
| Decoding speed | 15x faster |
| Context window | 1M tokens |
The prior gen in question is the M2.x SparseMoE architecture. At 1M tokens, standard attention has a quadratic scaling problem — compute and memory requirements blow up. MiniMax’s claim is that MSA solves this through selective attention, processing only the most relevant token relationships rather than computing all-pairs attention at full scale.
MiniMax has not released the full technical paper as of launch, so the exact mechanism of MSA is not independently verifiable yet. What matters for builders right now is the practical claim: at 1M token context, M3 is supposed to be fast and affordable to run, not just theoretically capable of handling long inputs.
Benchmark Position
MiniMax claims 59.0% on SWE-Bench Pro — the harder variant of the coding benchmark that tests against professional-grade software engineering tasks.
For comparison:
- Claude Opus 4.6: ~80.8% SWE-Bench Verified (different benchmark variant)
- MiniMax M2.7: 56.22% SWE-Bench Pro; ~78% SWE-Bench Verified
- MiniMax M2.5: 80.2% SWE-Bench Verified
The 59.0% represents a step up from M2.7’s 56.22% on the SWE-Bench Pro variant. On SWE-Bench Verified specifically, M2.7 actually regressed to ~78% compared to M2.5’s 80.2% — making M3’s improvement directionally important, even though the two SWE-Bench Pro figures above (56.22%, 59.0%) are a different test than the two SWE-Bench Verified figures (78%, 80.2%).
Caveats that matter:
MiniMax ran their benchmarks on their own infrastructure with agent scaffolding. This is a known confounder in agentic coding benchmarks: the scaffolding, retry logic, and tool-calling setup can materially affect scores. Independent third-party reproduction is pending. The 59.0% should be treated as a claim to verify, not a confirmed number.
MiniMax also claims M3 beats GPT-5.5 and Gemini 3.1 Pro on coding benchmarks. These claims require independent verification. GPT-5.5 has been independently benchmarked extensively; if M3 does outperform it on coding tasks, that will show up in the Artificial Analysis and LMSYS Chatbot Arena data within a few weeks of launch.
Why M2.x History Matters for M3 Evaluation
MiniMax’s model quality has been real, but their release practices have been inconsistent in ways that affect builder trust:
M2.5 (February 2026): Strong 80.2% SWE-Bench Verified, Modified MIT license, genuinely competitive pricing at $0.15/M input. Well-received.
M2.7 (March 18, 2026 API launch, April 2026 weights): Self-evolution training claims, but SWE-Bench Verified regressed to ~78%. License shifted from Modified MIT to commercial-authorization-required — meaning commercial use requires written permission from MiniMax. Community called it “faux open-source”. Input price doubled from $0.15 to $0.30/M.
M3 launches into this context. Builders who built pipelines on M2.5 expecting MIT-style openness were burned by M2.7’s license change. The license terms for M3 have not been announced as of this writing.
The License Question
This is the critical unknown at launch.
MiniMax has not published M3’s license terms. The progression from M2.5 (Modified MIT) to M2.7 (commercial-authorization-required) created real friction for teams that had committed to open-weight flexibility. Before building on M3 for commercial use, verify the license on the HuggingFace model card when weights release.
Specifically, check:
- Is commercial use permitted without written authorization from MiniMax?
- Are there restrictions on fine-tuning and redistribution?
- Does it permit deployment via third-party inference providers?
If M3 ships with terms similar to M2.7, teams that need genuine open-weight flexibility for self-hosted commercial deployment should wait for independent assessment before committing to the stack.
Access Paths
OpenRouter (available now)
M3 is live on OpenRouter with promotional pricing at approximately $0.30/M input tokens. This is OpenRouter’s 50% promotional rate — standard pricing will be higher once the promo period ends (see the June 14 update above: it has, to $0.60/M). OpenRouter provides an OpenAI-compatible API endpoint, so integration into existing pipelines requires only changing the base URL and model name.
For builders who want to evaluate M3 quickly without infrastructure overhead, OpenRouter is the starting point.
HuggingFace (weights releasing in ~10 days from June 1)
MiniMax has committed to releasing weights within approximately 10 days of API launch — targeting around June 10-11, 2026. Self-hosting requires substantial GPU infrastructure: M3’s parameter count was not disclosed at launch, but the technical report and HuggingFace model card later confirmed roughly 428B total parameters with ~23B active per token, up from M2.x’s 229.9B/9.8B. The MSA architecture’s efficiency claims apply most clearly at the 1M token context range — at standard context lengths, the architectural advantage is less pronounced.
build.nvidia.com / NIM — Not confirmed for M3. NVIDIA’s NIM catalog includes third-party models from various providers; whether M3 appears there post-launch depends on MiniMax’s partnership decisions.
Competitive Positioning
| Model | SWE-Bench | Context | Input Cost | Open Weights | Standout |
|---|---|---|---|---|---|
| Claude Opus 4.8 | 88.6% Verified | 1M | $5/M | No | Best coding agents, deepest reasoning |
| MiniMax M3 | 80.5% Verified† | 1M | $0.60/M | Yes | Long-context efficiency, open weights |
| MiniMax M2.7 | 56.22% SWE-Pro | 200K | $0.30/M | Yes (restricted) | Self-evolution training |
| Gemini 3.5 Flash | Strong agentic | 1M | $1.50/M | No | 4x faster token generation than other frontier models, GA stability |
†This is MiniMax’s own figure from its June 11 technical report, not an independent third-party evaluation (see the June 14 update above). MiniMax’s own launch-day figure was 59.0% on SWE-Bench Pro (different benchmark variant, also self-reported).
Correction: an earlier version of this table listed Claude Opus 4.8 at 200K context and $15/M input — both wrong. Anthropic’s own pricing documentation confirms Opus 4.8 (and Opus 4.6) run the full 1M-token context window at the standard $5/M input rate; there’s no separate long-context surcharge.
The clearest M3 advantage is price, not context: Claude Opus 4.6 and 4.8 also ship a 1M-token context window at standard pricing (GA since March 2026), so M3 doesn’t have a context-size edge over current-generation Claude. What it has is a large price gap — Claude Opus’s $5/M input is roughly 8x M3’s $0.60/M standard rate (or about 17x M3’s now-expired $0.30/M launch promo). Against Gemini 3.5 Flash’s $1.50/M, M3’s promotional rate was 5x cheaper; at M3’s current $0.60/M standard rate the gap is about 2.5x. If MiniMax’s speed claims for MSA at 1M context hold up, M3 becomes a serious candidate for any pipeline that needs to process long documents, large codebases, or extended conversation histories at scale where budget, not context length, is the binding constraint.
Who Should Evaluate M3 Now
Strong candidates:
- Teams building RAG or agent pipelines on long documents — legal, research, codebase analysis — where 1M context and $0.30/M pricing changes the economics meaningfully
- Teams already using MiniMax M2.5 or M2.7 who want to see if M3’s improved SWE-Pro score and new architecture upgrade their pipeline performance
- Teams evaluating open-weight alternatives who are waiting for Nemotron 3 Ultra’s June 4 launch and want a comparison point
Should wait:
- Teams that need commercial-use clarity — hold until the HuggingFace model card confirms license terms
- Teams that need independently verified benchmarks — the 59.0% SWE-Pro claim needs third-party reproduction before it should influence production decisions
- Teams that need stable, managed infrastructure with SLAs — M3 is new and OpenRouter pricing is promotional; long-term cost projections are uncertain
What to Watch
June 1–11: Independent benchmarks from Artificial Analysis, LMSYS, and community evaluations. These will confirm or complicate MiniMax’s 59.0% SWE-Bench Pro claim.
~June 10–11: Weights release on HuggingFace. License terms will become visible at that point — check the model card before committing to M3 for commercial self-hosted use.
June 4: Nemotron 3 Ultra launches on the same date range. The two models compete on similar ground: open weights, 1M context, agent-optimized training. Benchmark comparisons between them will clarify where each has an edge.
We will update this guide when weights land and independent benchmarks are available.
We research models; we do not test or deploy them directly. Benchmark data in this article is sourced from MiniMax’s official announcement, OpenRouter’s model listing, and third-party coverage as of June 3, 2026.