MiniMax M3: An Open-Weight Model Targets the Frontier

On June 1, 2026, MiniMax released M3 and described it as the first open-weight model to combine frontier agentic coding, native multimodality, and a 1-million-token context window in a single system (MarkTechPost, Pandaily). It arrived a week before Anthropic opened its Mythos-class model to the public, and the two releases bracket where the open and closed frontiers stand right now.

If the reported numbers hold, M3 is among the most capable open models released this year. That conditional matters, so let's take the claims and the caveats together.

The Architecture: MiniMax Sparse Attention (MSA)

The main engineering story is MSA — MiniMax Sparse Attention — an attention design built to make long context cheap to serve (MarkTechPost):

• ~1/20th the per-token compute at 1M-token context versus the prior M2 generation
• >9x faster prefill and >15x faster decode at full context length
• >4x faster than Flash-Sparse-Attention

This is the part worth attention independent of the leaderboard. Million-token context has been gated by cost: standard attention compute grows quadratically with sequence length, so a model that advertises a 1M window often becomes too slow or too expensive to actually use near the top of that range. MSA is MiniMax's approach to keeping the window large while keeping it affordable to run. For an open-weight model that organizations self-host on their own hardware, inference cost is the deciding factor, so an architecture that cuts per-token compute by an order of magnitude at full context is arguably more consequential than a benchmark point or two.

Concretely, the M2-to-M3 comparison is the one to anchor on: MiniMax reports roughly a twentieth of the per-token compute at 1M context relative to its own previous generation, alongside the prefill and decode speedups above. If those hold up on released weights, the practical effect is that workloads which were technically possible but economically painful on earlier long-context models — full-repository code analysis, long document sets, extended agent trajectories — become routine rather than exceptional. That is the kind of efficiency claim that, if verified, matters more for adoption than a few points of benchmark headroom.

The model is also natively multimodal — image input, video input, and desktop computer-use are built in rather than added later, with training data described as on the order of 100 trillion tokens. MiniMax did not disclose the parameter count or the training hardware, which is the first of several specifics that will only be verifiable once the weights are public.

The Benchmarks (Read With the Caveat)

MiniMax's reported results place M3 in frontier-adjacent territory on coding and agentic tasks (MarkTechPost):

The SWE-bench Pro figure is the notable one: 59.0% on agentic coding, ahead of two closed frontier models from OpenAI and Google, in a model you can download and run yourself. The picture is not uniformly strong — PostTrainBench sits below both Opus 4.7 and GPT-5.5 — but the spread is what you'd expect from a model tuned heavily for agentic coding and tool use rather than every task type.

A few of the agentic results are worth reading alongside SWE-bench Pro, because they fill in the shape of the model. OSWorld-Verified at 70.06% measures real desktop computer-use over up to 200 steps, MCP Atlas at 74.2% covers tool and agent orchestration, and KernelBench Hard at 28.8% reflects low-level GPU kernel generation on Blackwell hardware — a narrow, difficult task where a high score would be notable and a low one is expected. Taken together they describe a model built for agentic, tool-heavy work rather than one optimized for a single leaderboard.

The caveat is important, and several outlets led with it. These are self-reported, unverified numbers, and neither the weights nor the technical report were available on release day (TechTimes). Self-reported coding benchmarks are particularly sensitive to the scaffolding around the model — the agent harness, the number of attempts, and the evaluation harness all move the headline number — and MiniMax noted its SWE-bench runs used Claude Code scaffolding averaged over four runs. That last detail matters: a result produced with a strong external agent harness measures the model-plus-harness, not the model alone, and the figure can shift when someone reproduces it with a different setup. Until independent parties can reproduce the results on released weights, the honest framing is "frontier claims, unverified benchmarks," and that is how we've treated the placement below.

Pricing and Token Plans

M3's pricing differentiates by call length — a long-context surcharge applies above 512K tokens — and the standard rates are low (OpenRouter, Kingy AI):

• Standard: $0.60 input / $2.40 output per 1M tokens
• Launch promotion (first week): $0.30 / $1.20
• Cache reads: $0.12 per 1M
• Long-context (512K–1M): 2x the standard rate

MiniMax also offers subscription token plans for its Code product: Plus at roughly 1.7B tokens/month for $20, Max at about 5.1B tokens/month for $50, and Ultra at around 9.8B tokens/month for $120 (MarkTechPost). Because the model is open-weight, the API price is a ceiling rather than a floor — self-hosting changes the math for high-volume workloads, and the long-context surcharge stops applying when you run the weights yourself.

The license wasn't finalized at launch. MiniMax has used a modified-MIT license for prior models, and the M3 terms will land with the weights (Kingy AI). If your adoption depends on commercial-use or redistribution terms, that's the detail to confirm when the repository goes live.

Where M3 Lands on the Leaderboard

We've added MiniMax M3 to the AI model leaderboard in the open-weight band, alongside DeepSeek V4 Pro and Kimi K2.6 — the models it most directly competes with. Given that the benchmarks were unverified at launch, we placed it conservatively rather than taking the SWE-bench Pro headline at face value:

For reference, the closed frontier sits above this band — Claude Fable 5 (100), Opus 4.8 (99), and GPT-5.5 (97) lead the overall table. M3's position isn't "beats the frontier" — it's a capable open, multimodal, long-context model at low cost. If post-release verification confirms the SWE-bench Pro number, the ranking will move up; if the released weights underperform the self-reported figures, it will move down. We'll adjust it either way once the weights are independently testable.

The Open-Weight Field It's Joining

M3 lands in an unusually crowded open-weight tier. Each of the major entrants is making a different bet, which is worth keeping in mind when you choose one:

• DeepSeek V4 Pro competes on price-to-quality, with an Apache 2.0 license that imposes the fewest restrictions of the group.
• Kimi K2.6 is tuned for long, multi-step agentic runs with high tool-call counts.
• GLM-5.1 leans toward reasoning, and its lineage notably trained on non-NVIDIA hardware.
• MiniMax M3 distinguishes itself on the combination of native multimodality and cheap-to-serve 1M context.

The through-line is that the open-weight gap has shifted from reasoning — where it largely closed over the past year — toward agentic coding and multimodality, the capabilities that tend to drive production deployments. That's a more demanding target, and the fact that multiple open models are now credibly attempting it is the real signal here, separate from any one launch's benchmark table.

Why "Open Weight" Changes the Calculation

When the weights ship, the practical differences from a closed model are concrete rather than ideological. You can run M3 inside your own infrastructure, including air-gapped or data-residency-constrained environments where sending tokens to a third-party API is not an option. You can fine-tune it on proprietary data and keep the resulting model private. You can inspect and quantize it to fit your hardware budget, and you are insulated from a vendor deprecating the model, changing its pricing, or altering its behavior underneath you. For long-context workloads in particular, self-hosting also sidesteps the 2x surcharge that applies to API calls above 512K tokens.

Those benefits are exactly why the launch-without-weights gap is worth flagging rather than glossing over. Until the weights and technical report are public, none of the above is actually available — what shipped on June 1 was API access and a set of self-reported numbers. The open-weight advantages, and the ability to independently verify the benchmarks, both arrive on the same day the repository does. That is the date to mark, not the announcement date.

Practical Guidance

The recommendation is unchanged from how we treat any new model: build a routing layer rather than committing to one model. Send routine traffic to a cheap, fast tier, escalate the hard tasks, and re-evaluate when a model like M3 ships verified weights. A model-routing approach lets you adopt M3 for the workloads where its price-to-capability wins — long-context document and codebase work, multimodal inputs, cost-sensitive agentic coding — without staking your whole stack on a benchmark table that wasn't reproducible on launch day.

If you're evaluating M3 specifically, the most informative test is your own: run it head-to-head against DeepSeek V4 Pro and your current default on a representative slice of your real workload, weighting long-context and multimodal cases where M3's architecture should show its advantage.

What to Watch When the Weights Drop

When the repository goes live — promised within about ten days of launch — a short checklist separates a model that lives up to its announcement from one that doesn't:

• Reproducible SWE-bench Pro. Does the 59.0% hold when an independent party runs it, and with what scaffolding? A figure that only appears under a specific agent harness is weaker than one that survives a neutral setup.
• The license terms. Modified-MIT is expected, but the specifics — commercial use, redistribution, any usage restrictions — determine whether you can actually deploy it the way you intend.
• Real serving cost at long context. MSA's compute claims are the most interesting part of the release. Verify the prefill and decode speedups on your own hardware near the top of the context window, where the architecture is supposed to pay off and where most long-context models fall down.
• Parameter count and hardware footprint. MiniMax didn't disclose model size. The number determines whether you can self-host it at all, and on what GPUs.

If those check out, M3's ranking moves up and it becomes a serious option for long-context and multimodal workloads. If they don't, it settles in as one more capable-but-overstated launch. Either way, the weights and technical report are the moment the claims become testable. When they land, we'll re-run the placement against independently verified scores and update this post.

---

See the live standings on the AI model leaderboard, and read our companion piece on the week's closed-frontier release, Claude Fable 5.

Sources: MarkTechPost · Pandaily · OpenRouter · Kingy AI · TechTimes — frontier claims, unverified benchmarks