MiniMax M3 — Open-Weight Frontier Model Beats GPT-5.5 on Code & Agentic Benchmarks — 2026-06-01

MiniMax M3 landed on June 1, 2026, with benchmark results that knock GPT-5.5 and Gemini 3.1 Pro off their coding perch. The weights are public, so anyone can run it. The standout data point: in a fully autonomous 24-hour optimization run, M3 took GPU utilization on an Nvidia Hopper chip from 7.6% to 71.3%, a 9.4x speedup, without a single human intervention. For a model that currently runs $0.30 per million input tokens at promo pricing, that's a hard number to wave away.

Background you need

The frontier AI market has, until recently, been a closed-weights affair at the top. OpenAI's GPT-5.5, Anthropic's Claude Opus 4.7, and Google's Gemini 3.1 Pro hold the upper tier of coding and reasoning benchmarks, and none of them release their weights. That leaves developers dependent on API access, subject to rate limits, pricing changes, and vendor availability. Open-weight alternatives have historically trailed by a real margin on the benchmarks pros care about most: SWE-bench for real software engineering tasks, agentic workflow evaluation, and long-context reasoning.

MiniMax is a Chinese AI lab with several years of model work behind it. M3 is a step change in how the company positions itself publicly. It's built to fight the best closed models head-to-head while shipping its weights for self-hosting. The timing is pointed. It arrives as the developer community grows wary of vendor lock-in and hunts for open alternatives that don't give up capability.

The architecture is called Minimax Sparse Attention, or MSA. Dense attention, used by most transformers, computes relationships between every token pair in a sequence. MSA processes only the most relevant slices of a long input. That efficiency is what lets M3 support a context window up to 1 million tokens, with a guaranteed minimum of 512,000. Most frontier models cap at 200,000 or fewer. A 1M window means M3 can hold entire large codebases, long research documents, or hours of conversation at once without truncation.

What's new

On SWE-bench Pro, the harder, more recent version of the standard SWE-bench engineering benchmark, M3 passes both GPT-5.5 and Gemini 3.1 Pro and lands just below Claude Opus 4.7. On SVG generation quality, it beats Opus 4.7 outright. The MCP Atlas benchmark, which tests how well a model uses tool-calling interfaces the way a developer assistant would, also went to M3. On Browser Comp and OS World, it beats both Opus 4.7 and GPT-5.5. Across PostTrain Bench, a composite ranking over many dimensions, M3 placed third overall, behind Opus 4.7 and GPT-5.5 only.

The CUDA kernel demo is worth describing in full, because it shows something a benchmark table can't. MiniMax handed M3 a non-runnable Triton skeleton, a rough starting point for a CUDA kernel targeting Nvidia's Hopper architecture, and stepped back. Over the next 24 hours the model ran 147 distinct optimization iterations and made roughly 2,000 tool calls, testing, measuring, and revising with no prompting or course-correction. The final kernel hit 71.3% GPU utilization, up from 7.6%. Our read: the bigger detail isn't the outcome, it's the reliability. Across thousands of sequential decisions, the model didn't hallucinate its way into a dead end. Long-horizon agentic reliability has been a weak point for most models. A 24-hour autonomous run that doesn't derail is a meaningful data point, even as a single controlled demo.

M3 is also natively multimodal. Vision and language were co-trained from the start, not bolted on after. A visual form-filling test, where the model was shown a blank form image and asked to place fields in their correct positions, showed genuine spatial reasoning. On front-end code generation, reviewers rated M3's UI output, including components, animations, and typography, above Qwen 3.7 Max. Its Three.js 3D scenes handled physics simulations and procedural graphics, and its SVG outputs ran past 2,000 lines with animation, scene composition, and day/night cycle rendering.

Pricing sits at $0.30 per million input tokens and $1.20 per million output tokens under a current 50% promo discount. Standard rates are $0.60 and $2.40. A $20 monthly plan yields roughly 1.7 billion M3 tokens. You can reach the model through MiniMax's API, the M Code platform, OpenRouter, and, as of this writing, for free on OpenCode. The open-weight release means developers with the hardware to host it can run M3 at zero per-token cost.

The pushback

The obvious caveat: benchmark leadership is provisional. GPT-5.5 and Gemini 3.1 Pro are actively maintained, and the gap M3 has opened on SWE-bench Pro and MCP Atlas could narrow fast with a single new release from either lab. The CUDA demo, impressive as it is, is one constructed benchmark with a defined start and a measurable success metric. Real production workloads are messier. And 24 hours of autonomous compute isn't free even at promo pricing. The cost of extended autonomous runs adds up faster than a per-token headline rate suggests.

Then there's enterprise trust. MiniMax is less established in Western developer ecosystems than OpenAI or Google. Open-weight availability is a genuine win for independent developers, but it also means there's no managed safety layer between the model and whoever deploys it. Organizations with strict compliance needs will have to evaluate M3's behavior themselves rather than lean on a vendor's audit infrastructure. The third-place PostTrain Bench finish is strong, and it also confirms M3 isn't uniformly dominant across every dimension. Opus 4.7 and GPT-5.5 still lead on certain axes. Knowing which axis matters for your use case takes testing, not leaderboard-reading.

Bottom line: if you're an independent developer or a team chasing vendor independence on coding and agentic work, M3 is the open-weight model to test first this cycle. If you're in a compliance-bound org leaning on a vendor's safety guarantees, the open weights cut both ways, and you'll want your own evaluation before it touches production.

Source