senn-techAI Radar: MiniMax M3 — Open-Weight with Million-Token Context
We already had MiniMax on our radar as an underestimated challenger. With the leap from the M2 series (M2.5, M2.7) to M3, the provider becomes significantly more interesting — especially for a use case often underestimated in the SME sector: very long contexts.
What Makes M3 Special
- One Million Token Context: entire files, contracts, or codebases fit into the model in one piece — without tedious chunking.
- New Attention (MSA): the "MiniMax Sparse Attention" filters relevant memory blocks beforehand, instead of comparing every token with every other. According to MiniMax, computational load drops drastically with full context, and inputs and outputs become many times faster. This is exactly what makes long contexts affordable in the first place.
- Natively Multimodal: text and image are processed jointly, not tacked on afterward.
- Coding & Agents: in initial public coding benchmarks (SWE-Bench Pro), M3 performs on par with leading proprietary models — interesting for agentic workflows.
- Open Weights: M3 appears on Hugging Face and GitHub, allowing it in principle to be self-hosted.
The Important Caveat
The hosted MiniMax API is operated from China. For personal or confidential data, this is a clear GDPR issue — the inputs leave the EU. And the million-token context is no free lunch: self-hosting requires serious GPU resources. Both need to be checked before productive use.
Our Take
We evaluate models by task, not by brand. M3 moves onto our test list — especially for context-heavy cases on self-hosted hardware, where data sovereignty and long documents converge. We pick the model that solves the task best.
Further Reading
- MiniMax-M3 on Hugging Face — open weights and model card
- MiniMax on Hugging Face — all M-Series models
- MiniMax (company) — Wikipedia
- SWE-bench leaderboard — independent coding benchmark
- MiniMax-M2.7 analysis — independent data on the previous generation
Is MiniMax M3 worth it for our self-hosting?+
Mainly for context-heavy tasks on your own hardware. M3 offers a one-million-token context and ships with open weights on Hugging Face and GitHub, so it can in principle be self-hosted. The new MiniMax Sparse Attention cuts compute at full context. For short tasks or pure API use, though, there are cheaper or more privacy-friendly alternatives.
What about privacy and GDPR with MiniMax?+
The hosted MiniMax API is operated from China — personal or confidential data leaves the EU, which is a clear GDPR issue. You only get data sovereignty by self-hosting the open weights, not by using the cloud API. Using the API alone swaps a US hyperscaler for a Chinese provider without solving the underlying problem.
What is special about M3’s new attention?+
The MiniMax Sparse Attention (MSA) filters relevant memory blocks beforehand instead of comparing every token with every other. According to MiniMax, compute drops drastically at full context, and inputs and outputs become many times faster. That is exactly what makes the million-token context affordable in the first place — otherwise the KV cache quickly becomes the bottleneck at this length.