MiniMax M3: 1M Context, Open-Weight, Agentic Frontier

MiniMax M3 launched on May 31, 2026, and the pitch is bold: the first open-weight release to hold frontier coding, a one-million-token context window, and native multimodality at the same time. The architecture behind it is MiniMax Sparse Attention, the launch price is a fraction of closed frontier, and the API was live day one. The honest framing matters just as much as the headline.

Two things temper the excitement, and a credible read has to surface both. First, every benchmark MiniMax published was run on its own infrastructure with no independent validation at launch. Second, the open weights were not actually available on launch day; MiniMax committed to releasing them within roughly ten days, and the license was unconfirmed. So the "fully open" story is, for now, a promise rather than a delivered fact.

This guide covers what shipped, the sparse-attention design that makes 1M context affordable, the vendor-stated benchmark picture, the awkward timing against Claude Opus 4.8, the pricing math that makes the token plans genuinely interesting, and a clear decision framework for who should adopt now versus wait.

What went live on May 31 was the model behind an API, not a weights drop. M3 appeared same-day on OpenRouter under minimax/minimax-m3, exposed a 1M-token context window, and shipped with day-one compatibility for IDE integrations including Claude Code, Cursor, Roo Code, and Cline. The API uses a toggleable thinking mode: on for deep reasoning and long-horizon planning, off for low-latency completion.

MiniMax positions M3 as a clean version break from the M2 line, not a decimal increment. Its direct predecessor was M2.7, a self-evolving model that the company reported was handling a meaningful share of its internal reinforcement-learning workflow autonomously. M3 carries that agentic-first philosophy forward and adds native multimodality. If you followed the lineage, our MiniMax M2.7 and MiniMax M2.5 guides set the context for how aggressively this team has been iterating.

One detail builders should not over-read: M3's total parameter count is undisclosed. The M2.7 predecessor was a 229B-total / 9.8B-active mixture-of-experts model, but MiniMax has not confirmed M3 inherits those figures, so we treat the size as unknown rather than carrying forward old numbers. Always read the official model card and license text once the weights actually publish.

The reason this release reads as ambitious is that those three properties have historically pulled against each other. Quadratic attention scaling makes a genuinely usable million-token window expensive, which is why long-context models have leaned on compression tricks that trade away precision. Multimodality bolted on after the fact has tended to weaken visual reasoning rather than strengthen it. And frontier-level coding has typically demanded the kind of dense compute budgets that push long-context inference costs past what most teams will pay.

MiniMax's framing is that M3 breaks all three constraints at once: a sparse-attention design that keeps long context affordable, a training corpus that was multimodal from step zero, and agentic coding scores it claims rival closed frontier. Whether the model fully delivers on that promise is exactly the question independent benchmarks have not yet answered. The architecture is real and documented; the capability claims are vendor-stated.

The technical centerpiece is MiniMax Sparse Attention (MSA). Rather than computing full quadratic self-attention across the whole sequence, MSA performs block-level selection on the key-value cache. The distinction MiniMax draws against latent-compression approaches matters: where some designs compress key-values into a smaller latent space, MSA selects relevant blocks of uncompressed grouped-query key-values. The claimed benefit is that precision and prefix-caching compatibility are preserved, because the model still attends to the actual stored representations rather than a lossy summary.

The kernel engineering is part of the story. MiniMax describes a "KV outer gather Q" pattern, where key-value blocks form the outer loop and every query that hits a given block is batched together so that block is read from memory exactly once, in contiguous rather than scattered access. MiniMax claims this runs more than four times faster than open-source sparse-attention alternatives such as Flash-Sparse-Attention or flash-moba. As with the rest of the architecture claims, that figure is vendor-stated.

Read the bars carefully. The compute figure is the model's claimed reduction to roughly one-twentieth of M2's per-token cost at one million tokens, and the speedups are the vendor's prefill and decode numbers on its own hardware. The earlier MiniMax teaser cited more precise figures of about 9.7 times prefill and 15.6 times decode; the final launch materials rounded these to more than nine and more than fifteen times. Either way, none of it has been reproduced by a third party, so treat it as a design claim worth testing rather than a settled result.

MiniMax published a strong agentic benchmark sheet. Before quoting any of it, set expectations: these scores were produced on MiniMax infrastructure and had no independent validation at launch. If you are unsure what these evaluations actually measure, our SWE-Bench Pro and Terminal-Bench 2.1 guide breaks down the methodology. The headline figures: SWE-Bench Pro 59.0%, Terminal-Bench 2.1 66.0%, OSWorld-Verified 70.06% for computer-use task completion, BrowseComp 83.5 for autonomous web search, and MCP Atlas 74.2% for tool use.

Two long-horizon autonomy demos give a more textured sense of what M3 can attempt. In one, MiniMax reports the model ran roughly twelve hours without human intervention, produced 18 commits and 23 experimental figures, and reproduced an ICLR 2025 award-winning paper with a vendor-stated reproduction score of 0.650. In another, M3 reportedly improved NVIDIA Hopper FP8 GEMM hardware utilization from 7.6% to 71.3% across 147 submissions over about a day with no reference solution, where comparable models gave up after a few dozen attempts. Both demos are vendor-reported.

That caveat is the editorial spine of this release. A useful contrast: other recent frontier launches had independent intelligence-index numbers within roughly a day. M3 did not, at launch. The disciplined move is to wait for independent leaderboard and arena results before treating any of these figures as production-grade evidence, then run your own evaluation on the prompts you actually care about.

Here is the framing most day-of coverage either missed or buried. M3 launched on May 31, three days after Claude Opus 4.8 shipped on May 28. MiniMax's comparisons were set against Claude Opus 4.7, the pre-Opus-4.8 frontier. On the agent benchmarks where a direct comparison exists, the newer Opus 4.8 leads M3 by double-digit margins: SWE-Bench Pro 69.2% versus 59.0%, Terminal-Bench 2.1 74.6% versus 66.0%, and OSWorld-Verified 83.4% versus 70.06%.

That is not a reason to dismiss M3 — it is a reason to frame it correctly. The accurate read is that M3 lands roughly level with Opus 4.7 on agentic work while costing a small fraction of the closed price, and trails the three-days-older Opus 4.8 by ten to fourteen points on directly comparable evals. Stated that way, the story stays compelling without overclaiming. For the closed-frontier reference point, see our Claude Opus 4.8 release coverage.

The cost story is where M3 makes its strongest case. Launch pricing is $0.30 per million input tokens and $1.20 per million output on a limited-time 50% promotion, with a standard rate of $0.60 and $2.40 after. At standard rates that is roughly 8% to 20% of leading closed-frontier per-token pricing. Requests up to 512K input tokens bill at the standard rate; longer contexts cost more, with the exact surcharge not publicly disclosed at launch. For broader context on how this sits against the rest of the market, see our API pricing comparison.

The subscription tiers are the more interesting wrinkle. MiniMax offers shared multimodal quota across text, image, speech, and music: a Plus tier at roughly 1.7 billion tokens a month with 3 to 4 concurrent agents, a Max tier at roughly 5.1 billion tokens with 4 to 5 concurrent agents, and an Ultra tier at roughly 9.8 billion tokens with 6 to 7 concurrent agents. For high-volume agentic builders the breakeven math is striking, which is what the table below works out.

The "open-weight" label is doing a lot of work in the launch messaging, so be precise about it. On May 31 the weights were not on Hugging Face, and the license was unconfirmed — the candidates named in coverage were a permissive open license, but nothing was settled. MiniMax committed to publishing the weights and a technical report within roughly ten days. For anyone planning on-prem deployment, fine-tuning, or sovereignty-bound use, that means the open story is a near-term commitment to verify, not a capability you could act on at launch.

There is also a governance consideration that belongs in any enterprise evaluation, even for API use. As a Chinese company, MiniMax operates under China's 2017 National Intelligence Law, which obligates domestic firms to support and cooperate with state intelligence work. That applies to API-routed prompts regardless of where the user sits. It is not a reason to rule M3 out, but teams handling sensitive or regulated data should account for it alongside the usual model-selection criteria.

M3 is not a universal default, but it is a strong fit for specific profiles today and a wait-and-watch for others. The deciding factors are usage volume, tolerance for unvalidated benchmarks, and whether you need open weights now or can act on the API while they ship.

For most agencies and engineering teams the right first step is a scoped evaluation: run M3 on the prompts and repositories you actually care about, measure token spend and latency against your current default, and decide per-workload rather than per-headline. If you want help structuring that comparison, our AI digital transformation engagements start with exactly this kind of model-selection eval, and our development team can wire the winning model into your agent stack.