MiniMax shipped M3 today, an API-first launch with open weights and technical report promised within 10 days. The headline is MSA (MiniMax Sparse Attention), a new sparse-attention scheme that delivers 1M-token context at roughly 1/20 the per-token compute of the previous M2 model. Native multimodality is in: text, image and video as input (text-only output), trained together from the start rather than bolted on with a vision adapter. Pricing starts at $20/month for ~1.7B tokens; SWE-Bench Pro lands 59.0% (MiniMax says ahead of GPT-5.5 and Gemini 3.1 Pro, approaching Opus 4.7).
MSA is a two-stage attention mechanism. A lightweight index branch scans incoming tokens and selects which KV-cache blocks are relevant; the expensive attention computation then runs only on those selected blocks. MiniMax describes the implementation as "KV outer gather Q": KV blocks serve as the outer loop, aggregating the queries that hit them, each block read once with contiguous memory access. They claim greater than 4x faster than Flash-Sparse-Attn implementations. At 1M context, MSA produces approximately 15.6x faster decoding and 9.7x faster prefill versus M2, for 1/20 the per-token compute. Parameter count is not disclosed in the launch material. Open weights and technical report are scheduled within 10 days; the HuggingFace repo to watch is huggingface.co/MiniMaxAI. Until then, treat the architecture description as MiniMax's own framing pending independent verification, and the benchmark numbers as vendor-led (TechTimes already framed them as "frontier claims, unverified benchmarks").
Two ecosystem threads worth tracking. First, the cost angle: MiniMax is pricing M3 at roughly 5 to 10 percent the cost of GPT-5.5 and Gemini 3.1 Pro for similar coding benchmarks. If the numbers survive independent reproduction, this is the open-weight pressure point coming into focus for the closed frontier. Second, the architectural thread: MSA is a learned router (the index branch) selecting which KV blocks to attend to. That is structurally similar to MoE routing, applied to the attention layer instead of the FFN. For builders thinking about long-context efficiency without quadratic blowup, the index-then-attend pattern is one of the directions worth studying alongside Hyena, Mamba and longformer variants. The agentic side (Agent Teams running concurrent multi-stage workflows, the producer-verifier self-correction loop) is interesting but vendor-described; the 1,959-tool-call CUDA-optimization demo and the 12-hour autonomous paper-reproduction run are the kind of demos that look amazing in launch material and need real harness rerunning before they tell you something about your workload.
Monday morning, if you have long-context workloads (codebase reads, full-document QA, multi-hour agent runs), M3 is worth integrating via API today and benchmarking against your current frontier model in the 200K-to-1M range. The cost gap is large enough to matter even if M3 is meaningfully behind on quality, and the 512K guaranteed-fidelity floor gives you a safe operating band. If you are committed to open weights, the 10-day countdown matters: until those weights land on HuggingFace, this is a closed-API story. If you are building your own sparse-attention infrastructure, the MSA paper (when it drops) is required reading. And as always with vendor benchmarks against frontier models, run your own harness on a workload that matches yours before reorienting your stack.