Moonshot Kimi K3.1 Pro Nears Opus 5.5, Open Weights Nov 4
2 min readBeijing-based Moonshot AI released Kimi K3.1 Pro on October 23, a 2.8 trillion parameter mixture-of-experts model that scores within a point of Claude Opus 5.5 on Terminal-Bench 4.0 and beats it on SWE-Marathon. Moonshot says the weights will be open by November 4, which would make it one of the largest open-weight models ever shipped.
Background
Moonshot is one of China’s “six AI tigers,” and its Kimi line has been the strongest open-weight challenger to US labs on agentic coding. K3.1 Pro keeps the same architecture as K3, with 104 billion active parameters per token and a 1 million token context window, and it arrives just two days after rival DeepSeek closed a $15 billion funding round.
The numbers
On Moonshot’s own figures, measured at maximum reasoning effort, K3.1 Pro scores 65.7 percent on Terminal-Bench 4.0 against 66.4 for Opus 5.5 and 58.2 for GPT-6 Astra. It leads on SWE-Marathon at 51.7 versus 46.3 for Opus 5.5, and claims state of the art on OfficeQA Pro v2. Science is the weak spot: 38.1 on Terminal-Bench Science against 68.1 for GPT-6 Astra. Moonshot also says its claim-level hallucination rate fell from 23.4 percent on K3 to 8.3 percent. None of these results have been independently replicated yet.
Pricing drops to $2 per million input tokens and $10 per million output, down from $3 and $15 on K3. A Highspeed variant runs two to three times faster at double the price. The model is live on Kimi.ai, Kimi Code and an OpenAI-compatible API.
A deliberately blunted hacker
The most unusual detail is what Moonshot removed. The company says it stripped cyber-related data from mid-training and reinforcement learning before the open release. On its internal xdev-bench-hard test, an unfiltered checkpoint solved 90.5 percent of tasks while the shipped model solves 81 percent. Moonshot frames this as a trade: a safer open model in exchange for some vulnerability research skill, and it says K3.1 Pro still trails frontier proprietary models on offensive cyber tasks.
Why it matters
China’s open-weight labs are now shipping models that match the closed US frontier on coding at a fraction of the price, and Moonshot is doing it while publicly dialing back dual-use capability, a direct answer to the kind of concerns raised this week by Anthropic’s own cyber incident report. Watch for the November 4 weights drop, the license terms, and whether independent benchmarks confirm the coding numbers.