MiniMax M3: Open-Weight Frontier Model With Sparse Attention
Shanghai-based AI lab MiniMax unveiled its new language model MiniMax M3 on June 1, 2026 — and the bet here is not on more parameters but on a new architecture. At the center sits MiniMax Sparse Attention (MSA), a method designed to slash the per-token compute cost of long context. M3 is available via the API, monthly token plans and the agent product „MiniMax Code”. MiniMax positions the model as the first open-weight system to combine frontier-level coding, a one-million-token context window and native multimodality in one. The open weights themselves, however, are not yet out at launch — and several of the benchmark numbers cited come from the vendor itself.
What actually changed
- A new architecture instead of a bigger model: MiniMax Sparse Attention (MSA) cuts per-token compute at 1M context to 1/20 of the previous generation (M2), per the vendor — with more than 9x faster prefill and more than 15x faster decoding.
- 1M token context plus native multimodality (image and video) in a model that also targets agentic coding.
- An open-weight promise with a delay: model weights and the technical report are only set to ship „within ~10 days” of launch — they were not available on June 1.
- Pricing far below the closed frontier models: launch at $0.60 / million input tokens and $2.40 / million output tokens on OpenRouter, with a time-limited 50% promo bringing it to roughly $0.30 / million input and $1.20 / million output.
What held before
Most model releases over the past months followed one pattern: more parameters, more training data, more compute. Long context windows did exist — one million tokens has not been a differentiator since Gemini and others — but they were expensive to run. Classic attention scales unfavorably with context length: the more tokens in the window, the more compute each new token costs. That makes long context slow and pricey in practice.
At the same time, frontier coding and agentic work ran almost entirely through closed providers — OpenAI, Anthropic, Google. Anyone wanting top coding performance paid their API prices and could not run the weights themselves. The open world (DeepSeek, Qwen, Mistral) was catching up, but typically trailed a bit on the hardest coding benchmarks.
What holds now
With M3, MiniMax moves the marker in three places at once:
1. Sparse attention lowers the cost of long context. The core of M3 is MiniMax Sparse Attention. Instead of computing every token against all previous ones, MSA evaluates only a subset of the connections. According to MiniMax, this drops per-token compute at 1M context to one twentieth of the M2 generation, makes prefill more than 9x and decoding more than 15x faster. That is the real lever: not a bigger model, but a cheaper one to run at the same context length. It can only be verified, though, once the technical report and independent tests are out.
2. Frontier coding benchmarks — per the vendor. MiniMax cites, among others, 59.0% on SWE-Bench Pro, 83.5 on BrowseComp, 66.0% on Terminal-Bench 2.1 and 34.8% on SWE-fficiency for M3. On coding it claims to beat GPT-5.5 and Gemini 3.1 Pro and to approach Claude Opus 4.7. Important: these numbers are vendor-run, on its own infrastructure and against baselines it chose. Independent third-party scores were still pending at launch.
3. Open weights and an aggressive price. M3 launches on OpenRouter at $0.60 / million input tokens and $2.40 / million output tokens; a time-limited 50% promo pushes that to roughly $0.30 / million input and $1.20 / million output. Subscriptions start, per the vendor, at $20 / month. VentureBeat calculates that this puts M3 at roughly 5–10% of the cost of GPT-5.5 and Gemini 3.1 Pro. The model weights and technical report were announced for „within ~10 days” but were not yet available on launch day.
Context
Three things about this release are notable — and none of them is the raw benchmark number.
First, the architecture angle. If MSA holds up to MiniMax’s claims, M3 is less „the next, bigger model” than evidence that the cost of long context is a solvable efficiency problem. That would matter precisely for agentic workflows, which carry a lot of context anyway. But the claim only becomes solid with the technical report.
Second, the gap between announcement and availability. As of June 1, „open weight” is a promise, not a verifiable fact — weights and report are still missing. Until they land, M3 is effectively a closed API model with an open-source label. That is not an accusation, but an important distinction for anyone planning to build on the open weights.
Third, the necessary skepticism toward vendor benchmarks. TechTimes points out that every cited number comes from the vendor itself — on its own infrastructure, with baselines it picked. And the comparison ran against Claude Opus 4.7: the newer Opus 4.8 is back clearly ahead of M3 at 69.2% on SWE-Bench Pro. So „frontier” here is relative to the chosen cut-off date. Research also regularly shows a noticeable gap between lab benchmark and real-world deployment.
In the bigger picture, M3 joins the China open-weights wave: DeepSeek, Qwen and now MiniMax push on price and openness at the same time. Competitive pressure shifts from „who has the biggest model” to „who runs comparable performance most cheaply” — and that is exactly where the sparse-attention bet lands.
What you can do now
If you want to test M3 for coding agents: try it first via the API or OpenRouter at the promo price, rather than waiting for the open weights. That gives you your own read on real-world performance — independent of the vendor benchmarks.
If you depend on the open weights: wait for the announced Hugging Face / GitHub release and the technical report before committing to M3. Until then, the open-weight pledge is not yet fulfilled.
If you are comparing model costs: don’t treat the promo prices as permanent — calculate with the list prices ($0.60 / $2.40). Even then M3 sits well below GPT-5.5 and Gemini 3.1 Pro — but the 5–10% figure strictly only holds under the time-limited promo.
If you want to know who builds which model family and what it’s good for: → AI model families at a glance
Entdecke mehr
Usage-based instead of flat-rate: The AI cost turn was predictable — and here is how to absorb it
AI providers are pushing flat rates toward metered billing. Why it had to happen, what it costs and three levers that soften the shift.
GlossarMiniMax
MiniMax is an AI lab founded in Shanghai in 2021 whose open-weight M model family (LLMs) targets very long context windows and low running costs via sparse attention.
LexikonAI Pricing Made Simple
How AI models bill — tokens, input vs. output, hidden cost drivers and three levers to save. With price table and worked examples.