MiniMax M3 Tops Open-Weight Coding Benchmarks With a Million-Token Context — and a Pitch That No Government Can Recall It
When a US government order forced Anthropic to pull all public access to Claude Fable 5 and Mythos 5 earlier this month, it delivered an unintentional gift to a Chinese AI lab. MiniMax had launched its flagship open-weight model, M3, on June 1 with a claim that read like marketing boilerplate at the time: that weights you can download and run yourself cannot be switched off by anyone — not a vendor, not a regulator, not a state. Two weeks later, after watching two frontier models vanish from the internet on a national-security directive, that pitch looks less like boilerplate and more like a roadmap for how enterprises hedge their bets.
A frontier coding model you can host yourself
MiniMax, the Shanghai-based lab, bills M3 as the first open-weight model to combine three capabilities that had been the exclusive territory of proprietary systems: frontier-level coding-agent performance, a one-million-token context window, and native multimodality in a single architecture. On SWE-Bench Pro, an established software-engineering benchmark, MiniMax reports M3 scoring 59.0% — ahead of OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro, and just shy of Anthropic's Claude Opus 4.7. The company also reports 66.0% on Terminal-Bench 2.1, 74.2% on MCP Atlas, and 83.5 on the BrowseComp autonomous-search benchmark, where it actually edges past Opus 4.7.
Those numbers carry an important asterisk. As The Decoder and TechTimes both noted, the benchmarks are vendor-run — several were measured on MiniMax's internal infrastructure using Claude Code scaffolding — and at launch the weights had not yet shipped, leaving the headline claims unverified by independent parties. MiniMax committed to publishing the weights and a technical report on Hugging Face and GitHub within roughly ten days of the June 1 launch, a window that has now closed; independent replication is the next shoe to drop.
What is harder to dismiss is the architecture. M3's million-token context rides on a new attention scheme MiniMax calls MiniMax Sparse Attention (MSA). Instead of comparing every token against every other token — the quadratic cost that makes long context ruinously expensive — MSA splits the key-value cache into blocks, runs a lightweight index branch to pick which blocks are relevant to a query, and sends only those into the full computation. MiniMax says the result is one-twentieth the per-token compute of its predecessor at one million tokens, with prefill more than nine times faster and decoding more than fifteen times faster. The company reports its operator-level implementation runs more than four times faster than open-source alternatives such as Flash-Sparse-Attention.
MiniMax also leaned hard on autonomy demonstrations rather than static scores. In one internal test, M3 reproduced an ICLR 2025 award-winning paper on LLM fine-tuning, running for nearly twelve hours unattended and producing 18 commits and 23 figures. In another, it spent roughly 24 hours optimizing an FP8 matrix-multiplication kernel on Nvidia Hopper GPUs — 147 benchmark submissions, 1,959 tool calls — and pushed hardware utilization from 7.6% to 71.3%, with its best solution not arriving until submission 145. MiniMax noted that most other models stopped improving within the first 30 attempts; only Opus 4.7 and M3 kept going.
The "uncrecallable weights" argument
The strategic backdrop is what gives M3 its moment. According to VentureBeat, Anthropic blocked all public access to Claude Fable 5 and Mythos 5 following a US government order, with reporting elsewhere tying the shutdown to an export-control directive citing national-security authorities. All other Anthropic models, including Opus 4.8, remained available — but two frontier systems enterprises had been building on simply went dark, on a timeline and for reasons the customers did not control.
That is precisely the failure mode open weights are immune to. A closed model lives behind an API the vendor operates; if the vendor pulls it — whether for safety policy, a pricing change, a business pivot, or a government order — the capability evaporates and any workflow depending on it breaks. Downloaded weights cannot be recalled. Once an enterprise has M3 on its own hardware, no upstream decision can revoke it, and the data never leaves the building. For regulated industries, air-gapped environments, and buyers newly spooked by the Fable 5 episode, that durability is the whole sales pitch — and it lands harder in mid-June 2026 than it would have on June 1.
The trade-offs are real. Self-hosting a million-token, multimodal frontier model is not free: it demands serious GPU capacity, operational expertise, and a willingness to forgo the managed-service polish that closed vendors provide. M3 also still trails the very top of the closed field — Opus 4.8, which Anthropic shipped after M3's launch, is a stronger model on most benchmarks, and M3 landed behind both Opus 4.7 and GPT-5.5 on MiniMax's own PostTrainBench autonomous-training test. Open weights buy control and continuity, not necessarily the best raw scores. But for a growing class of buyers, control is the feature that matters most.
What to watch
The immediate question is verification. Now that the weights are out, independent benchmarking on SWE-Bench Pro and the long-context claims will either confirm M3's frontier positioning or knock the vendor numbers down a peg, as has happened repeatedly across the industry. Watch, too, for how quickly inference providers and self-hosters stand M3 up in production, and whether the MSA architecture delivers its promised efficiency outside MiniMax's own stack.
The larger story is geopolitical and structural. The Fable 5 shutdown showed that closed frontier models now carry a regulatory-recall risk that didn't visibly exist a year ago, and MiniMax — a Chinese lab — is positioned to absorb the customers that risk shakes loose, even as the same export-control dynamics that grounded Fable 5 could cut the other way. The open-versus-closed debate has been litigated mostly on cost and capability. M3 reframes it around something colder: who, exactly, gets to turn your AI off.