Kimi K3.1 Pro Benchmarks Land: Open Weights Hit No. 4
By @sharedot · · 8 pages
- AI Frontier
- Open Weight Models
- Kimi K3 1 Pro
- Benchmarks
- Moonshot AI
Independent tests score Moonshot's Kimi K3.1 Pro at 55 on the Intelligence Index, making it the strongest open-weight model in the ranking.
What Happened
Moonshot AI's Kimi K3.1 Pro, released October 23, 2026, has moved from announcement to independent verification. According to heise online, Artificial Analysis scored the model at 55 on its Intelligence Index, ranking fourth overall and up 11 points from the Kimi K3 score of 44.
With its weights scheduled for release by November 4, 2026, the model would currently be the strongest open-weight system in that ranking — a concrete benchmark milestone rather than a marketing claim.
Why It's Surprising
The surprise is how much performance Moonshot extracted without any new architecture.
Basic Tutorials notes the model is a retraining of Kimi K3's 2.8-trillion-parameter Mixture-of-Experts design (104 billion active parameters), with the gains coming from extensive reinforcement-learning post-training across hundreds of task domains.
Yet it now sits near Claude Opus 5.5 (58) and Claude Sonnet 5.5 (56) on the composite index — at a token price Basic Tutorials reports as half of Opus 5.5's.
The Evidence
The agentic numbers are the sharpest evidence. heise online reports that in Terminal-Bench 4.0 the success rate jumped from 13% to 67%, the best result in Artificial Analysis' tests, and AutomationBench-AA rule-compliant task completion rose from 58% to 73%.
In GDPval-AA, testing professional tasks, the model scored 1,846 Elo, ranking third behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857). Knowledge gains were modest — Humanity's Last Exam rose only from 47% to 52% — while the AA-Omniscience hallucination rate fell from 53% to 44%.
The Cyber Trade-Off
The benchmark rise came alongside a deliberate capability cut. National Technology News reports Moonshot removed cyber-related data from all training stages because, without that step, downstream cyber capability would increase substantially.
The company says the release model solved nearly 10% fewer problems on the xdev-bench-hard vulnerability benchmark than the pre-removal version, scoring below GPT-5.6 Sol and DeepSeek's V4.1-Flash, and that it aims not to release open weights that would extend the absolute frontier of cyber capability.
The Stakes for Builders
heise online reports task economics stayed roughly flat despite a third-lower token price: the model generated about 55,000 output tokens per Intelligence Index task, 14% more than K3, making an average task cost $2.11 versus $2.00 for K3 and $5.98 for Claude Opus 5.5 at maximum reasoning effort.
Standard throughput runs about 74 tokens per second, with the Highspeed variant at roughly 129 tokens per second — twice the token price but two to three times faster. For teams weighing open-weight deployment, those are usable numbers.
What Comes Next
The open-weight release is scheduled by November 4, 2026, which will let teams run the roughly one-million-token context model on their own hardware and probe the cyber-data removal independently.
Basic Tutorials cautions that most benchmark figures so far come primarily from Moonshot itself, with confirmation from independent labs such as Artificial Analysis and LMArena expected in the coming days given API availability.
Whether the stripped safeguards hold up once anyone can fine-tune the weights is the open question this release will actually answer.