Skip to main content
FeaturedDaily
Back to all news

Open-weight models

MiniMax ships M3 — open-weight frontier claims, weights to follow

Cheap, capable, and not quite downloadable yet. The brief on what shipped and what to watch.

By , Editor-in-Chief · FeaturedDailyVerified June 2026

The answer

MiniMax launched M3 on 1 June 2026, an open-weight frontier coding model with 1M context.

What happened

On 1 June 2026, Shanghai's MiniMax released M3, billed as the first open-weight model to combine frontier coding, a 1M-token context window and native image/video understanding in a single architecture. The underlying mechanism is MiniMax Sparse Attention (MSA), designed to cut prefill and decoding latency at long contexts — the practical bottleneck for agentic coding tasks. MiniMax reports 59.0% on SWE-Bench Pro, ahead of GPT-5.5 and Gemini 3.1 Pro on that test. Those figures are MiniMax's own.

The money. Launch pricing on OpenRouter: roughly $0.30 input / $1.20 output per million tokens (promotional). That's approximately a tenth of comparable closed-frontier rates — the part of this story that's verifiable on day one, because you can call the API and see the bill.

The catch and what to watch

The weights were not published at launch. MiniMax promised them on Hugging Face within approximately ten days of 1 June. Until they land and independent engineers reproduce the SWE-Bench numbers, the benchmark claims are vendor-reported and the 'open-weight' label is a commitment rather than a fact. Separately, ARC-AGI-2 sits under 12% — a real gap on abstract reasoning that MiniMax reported quietly. Strong on coding and long-context retrieval; weaker on flexible reasoning.

MiniMax describes it as the first open-weight model to combine frontier coding performance, million-token context, and native multimodal capabilities, though independent verification of those claims requires the model weights, which had not been released at launch.

Source: Tech Times · 1 June 2026

Frequently asked questions

What is MiniMax M3 in one line?
A Shanghai-built AI model launched 1 June 2026, pitched as the first open-weight system to combine frontier coding, 1M-token context and native multimodality — at roughly a tenth of the price of Opus or GPT-5.5.
Are the weights out?
Not at launch — MiniMax said they'd reach Hugging Face within about ten days. Until then the benchmark claims are vendor-reported and not independently verifiable.
What is M3 weak at?
Abstract reasoning: MiniMax reports under 12% on ARC-AGI-2, a real gap versus US frontier models. M3 is designed for coding and long-context tasks, not open-ended reasoning.

Sources

← All news