MiniMax-M3: The New Open Weights Frontrunner — Multimodal, 1M Context, Competitive Pricing
MiniMax launched M3, its most capable multimodal model: it leads the open weights rankings with 55 points on the Intelligence Index, supports image and video input, and offers a 1M token context window at competitive prices.
The open weights landscape keeps moving
MiniMax has released M3, and according to Artificial Analysis it scores 55 on the Intelligence Index — making it the top-ranked open weights model once the weights are publicly released (MiniMax has committed to doing so within ~10 days).
What M3 brings over its predecessor
M3 is the first model in the M-series with native multimodal input (text, image, and video), a significant leap over the text-only M2.7 (50 points).
Key improvements:
- Expanded context window: 1M tokens (up from 200K in M2.7)
- Stronger reasoning: +9 pts on HLE (37%), +6 pts on GPQA Diamond (93%)
- Instruction following: +7 pts on IFBench (83%)
- Competitive vision: ~80% on MMMU-Pro, on par with GPT-5.5
- Real-world tasks: ~1,670 on GDPval-AA (44 occupations, 9 industries), comparable to Claude Sonnet 4.6
Pricing:
- $0.30 / $1.20 per 1M input/output tokens (up to 512K context)
- $0.60 / $2.40 per 1M input/output tokens (512K–1M context)
One caveat worth noting:
On the AA-Omniscience benchmark, M3 only attempts 30.9% of questions — the lowest among current peers. This reduces hallucinations (16.1%) but also effective accuracy (15.0%). For use cases that require complete answers to ambiguous queries, this is a meaningful trade-off.
What this means for NapsixB2B customers
At NapsixB2B, we track the open weights space closely because it directly feeds our smart routing philosophy: having powerful, open, low-cost models available lets us assign the right task to the right model — no over-engineering, no wasted tokens on frontier models when they're not needed.
- If M3 delivers after weights are released, we'll evaluate its integration into the catalog for tasks requiring long context (1M tokens), native vision, or large-document processing.
- The competitive pricing ($0.30/$1.20) makes it attractive for high-volume agentic workflows.
- Current availability is through MiniMax's own API, SiliconFlow, GMI, and Novita — not yet on Bedrock or major hyperscalers.
Our technical team monitors benchmarks and license terms for every open weights release before enabling models on the platform. MiniMax's previous license (M2.7) carried commercial restrictions — we'll be watching M3's terms closely.
📊 Source: MiniMax-M3 — Artificial Analysis
