MiniMax's M3 Beats Every Open-Weight Rival at One-Tenth the Price
Artificial Analysis benchmarks MiniMax-M3 at 55 on its Intelligence Index — set to be the top open-weights model once weights ship, which hasn't happened yet
Takeaways
- Benchmark result: Artificial Analysis independently scores MiniMax-M3 at 55 on its Intelligence Index, making it the projected top open-weights model once weights ship.
- Weights not yet released: M3 is API-only right now; MiniMax committed to releasing weights on HuggingFace and GitHub within ~10 days of the June 1 launch.
- Key improvements over M2.7: HLE +9pts (28%→37%), GPQA Diamond +6pts (87%→93%), IFBench +7pts (76%→83%), with a 5x larger context window (200K→1M tokens).
- Architecture: MiniMax Sparse Attention (MSA) cuts per-token compute to 1/20th of the prior generation at 1M context, with 9x prefill and 15x decode speedups.
- Pricing: $0.30/$1.20 per 1M input/output tokens at launch promo (standard: $0.60/$2.40) — roughly one-tenth the cost of Claude Opus 4.7.
- Caveats: Benchmark scores are independently verified by Artificial Analysis, but self-reported coding benchmarks remain unverified; license terms and open-weights status are still pending.
Independent evaluation firm Artificial Analysis has published its full benchmark results for MiniMax-M3, scoring it 55 on the Artificial Analysis Intelligence Index. That number matters for one specific reason: once MiniMax releases the model weights , promised within roughly 10 days of the June 1 API launch , M3 will be the highest-scoring open-weights model on the index, edging ahead of Kimi K2.6 and MiMo-V2.5-Pro, both sitting at 54.
The distinction between "API available" and "weights available" is not a footnote here. Until the weights ship and independent engineers can inspect the architecture, verify the training setup, and assess safety behavior, M3's open-weight designation is a company commitment, not a verifiable fact. The benchmark results from Artificial Analysis are independently run, but the model itself is still a black box.
What the numbers actually say
The Artificial Analysis Intelligence Index v4.0 is a composite of 10 evaluations: GDPval-AA, τ²-Bench Telecom, Terminal-Bench Hard, SciCode, AA-LCR, AA-Omniscience, IFBench, Humanity's Last Exam, GPQA Diamond, and CritPt. It is run independently, not self-reported by the lab. Claude Opus 4.8 currently leads the index at 61, followed by GPT-5.5 at 60. M3's score of 55 puts it meaningfully behind the closed-source frontier but ahead of every open-weights peer.
The per-benchmark breakdown from Artificial Analysis shows real improvement over M2.7 (which scored 50):
- HLE (Humanity's Last Exam): +9 points, from 28% to 37%
- GPQA Diamond (scientific reasoning): +6 points, from 87% to 93%
- IFBench (instruction following): +7 points, from 76% to 83%
- AA-LCR (long-context reasoning): +5 points, from 69% to 74%
- SciCode (coding): small regression, from 47% to 45%
On GDPval-AA , a benchmark measuring real-world task performance across 44 occupations and 9 industries , M3 scores approximately 1670, behind Claude Opus 4.8 (1890) and GPT-5.5 (1769), but level with Claude Sonnet 4.6 at 1676. That is a meaningful result: on practical work tasks, M3 is roughly competitive with Sonnet-class closed models.
The multimodal upgrade
MiniMax positions M3 as the first domestic AI model to integrate frontier coding, agentic capabilities, million-token context windows, and native multimodal processing in a single architecture. The key word is "native" , M3 was trained on multimodal data from the ground up, with MiniMax rebuilding its entire data pipeline and scaling pre-training data to hundreds of terabytes to achieve tight alignment between text and visual semantic spaces. This is different from bolting a vision adapter onto an existing text model.
On MMMU-Pro (visual reasoning), M3 scores around 80%, level with GPT-5.5 at 79.9% and Kimi K2.6 at 79.4%, though behind Gemini 3.5 Flash at 84.3%. Not all open-weights models support native vision input, so this is a genuine differentiator in the open-weights space.
The abstention problem
One result stands out as a red flag for production use. On AA-Omniscience , a knowledge and hallucination benchmark , M3 attempts only 30.9% of questions, the lowest among current peers. This yields a low hallucination rate of 16.1%, but also a low accuracy of 15.0%. In other words, the model is staying quiet rather than being wrong, which is a conservative strategy that looks good on hallucination metrics but is not useful in practice. If you need a model that answers confidently across a broad knowledge domain, this is worth testing carefully before deploying.
The architecture behind the 1M context window
MiniMax Sparse Attention (MSA) replaces standard full-attention computation with a two-stage process: a lightweight index branch selects which blocks of the key-value cache are relevant to a given input, and the main attention layer then processes only those selected blocks. Standard attention scales quadratically with context length , double the context, roughly quadruple the compute. MSA breaks that relationship.
The practical impact is a 1M-token context window with per-token compute at 1/20th of the previous generation, plus a 9x prefill speedup and 15x decode speedup. This is notable because MiniMax itself abandoned sparse attention during the entire M2 generation in favor of full attention, calling the infrastructure "not yet mature" at the time. Bringing it back successfully at production scale is the real architectural story here.
Token efficiency is also worth noting. M3 uses approximately 91M output tokens to run the full Intelligence Index (about 81M of which are reasoning tokens), versus M2.7's 87M , scoring 5 points higher on the index for roughly the same token spend.
Where it fits in your stack
The practical use cases that benefit most from M3's combination of features:
- Long-horizon coding agents: MiniMax demonstrated M3 autonomously reproducing an ICLR 2025 outstanding paper, running for nearly 12 hours, producing 18 commits and 23 experimental figures without human intervention.
- Multimodal document analysis: A 1M-token context window can accommodate full annual reports, filings, and research archives in a single pass, with native multimodal support handling charts and diagrams.
- Large codebase review: The model processes up to 1 million tokens at once , five times more than M2.7 , allowing it to work across entire codebases and long-running agent sessions without losing track of earlier information.
- Browser and tool-use agents: On BrowseComp, a web search and browsing benchmark, M3 scores 83.5, surpassing Claude Opus 4.7's 79.3.
The honest positioning: M3 is competitive with closed frontier on coding and agentic browsing, ahead of every open-weight peer on long-context multimodality, and behind on raw abstract reasoning. Abstract, fluid reasoning is a known soft spot , the ARC Prize ARC-AGI-2 results from earlier this year had the MiniMax line scoring in the low single digits.
Pricing and access
The MiniMax direct API at platform.minimax.io uses tiered pricing by context length. During launch week, usage up to 512K context is priced at roughly $0.30 per million input tokens and $1.20 per million output tokens. OpenRouter lists the model as minimax/minimax-m3 at the same promotional rate, with standard pricing at $0.60 input and $2.40 output per million tokens.
For context,
Claude Opus 4.7 charges $5.00 per million input tokens and $25.00 per million output tokens
, M3's standard rate is roughly one-tenth of that.
For longer contexts, pricing steps up: input costs rise to $0.60/$2.40 for the 512K to 1M context tier. The model is available in two API versions , M3 and M3-highspeed , with identical results but faster inference on the latter. Automatic caching is supported and enabled by default.
The open-weights caveat
At the time of the API launch, neither the weights nor the technical report had been released. MiniMax said both would be made available within ten days, targeting publication on Hugging Face and GitHub for private cluster deployment and fine-tuning. The license terms are also unconfirmed. M2.7 shipped under a "Modified-MIT" license that blocked commercial use without written permission , M3 is expected to follow the same playbook. If self-hosting for commercial work is part of your plan, hold off until the actual terms are published.
There is also a data residency consideration that deserves explicit attention. MiniMax is a Chinese company, and China's 2017 National Intelligence Law requires organizations operating in China to cooperate with state intelligence efforts when requested. This is not unique to MiniMax, and it does not make the model unusable , but it is a real factor for anyone processing sensitive or regulated data through the hosted API. Self-hosting the weights, once available, changes that calculus substantially.
The Artificial Analysis evaluation gives M3 its first independent score, and the number is strong enough to make it the open-weights leader on the index. Whether that holds up under broader independent testing , and whether the weights actually land with workable license terms , is what the next two weeks will answer.