Kimi K3.1 Pro: The 3T Open-Weight That Catches Claude and GPT-6

Kimi released K3.1 Pro yesterday, their most capable 3T-class open-weight model to date, and it is turning heads. With a 1-million-token context window and benchmark numbers that match or approach the strongest proprietary models, this is the most competitive open-weight LLM release in months [1].

They also shipped K3.1 Pro Highspeed, a variant delivering 2-3x the speed of the base model for more responsive interactions. Both are available now on kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Full model weights will be released by November 4 [2].

The benchmarks

On Terminal-Bench 4.0, which evaluates coding agents on realistic terminal tasks with strict task-specific tests, K3.1 Pro scores 65.7%. That beats GPT-6 Astra (58.2%) and sits narrowly behind Claude Opus 5.5 (66.4%). On their internal Kimi Code Bench 3.0 Long-horizon, which tests multi-step execution and tool use across engineering domains, it scores 61.2%, up from K3's 55.95% [3].

For 3D creation, on BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, K3.1 Pro scores 92.7% with tools, approaching Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%) [4].

In finance, K3.1 Pro achieves state-of-the-art on OfficeQA Pro v2 (64.4), which tests complex fiscal data analysis over real financial documents through retrieval, cross-document verification, and multi-step numerical reasoning [5].

What is actually new

K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute. The team optimized every layer of the stack for large-scale joint reinforcement learning across hundreds of downstream domains: the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for environment quality checks [6].

The result is consistent gains in every domain they tested. Coding, agentic tasks, knowledge benchmarks, 3D modeling, financial research, all improved. The hallucination rate on their internal stress-test benchmark dropped from 23.4% (K3) to 8.3% (K3.1 Pro), and the share of responses containing at least one factual error fell to 42.4% of K3's rate [7].

The cybersecurity angle

Here is the interesting part. Kimi spent considerable effort understanding where cyber capability comes from in these models and how to control it. They found that cyber-related content was embedded in the general mid-training corpus and cyber-related tasks were present in the RL environment mix, a natural byproduct of training on broad data [8].

For the open-weight release, they identified and removed cyber-related data from both training stages while retaining general-purpose coding tasks. The impact is measurable: the unfiltered checkpoint solves 90.5% of their xdev-bench-hard vulnerability exploitation tasks at pass@3, while the final K3.1 Pro reaches 81.0%. That is a meaningful reduction, though the team notes that stronger general coding and debugging capabilities still generalize to cyber tasks even after targeted data removal [9].

Kimi states K3.1 Pro's offensive cyber capability remains behind frontier proprietary models that have been publicly available for at least three months. Their open-weight principle: aim not to release weights that extend the absolute frontier of cyber capability, while progressively raising the cyber capability available in open weights for defensive purposes [10].

Why this matters

The gap between open-weight and proprietary models has been closing steadily, but K3.1 Pro is one of the clearest examples yet of an open model matching frontier performance on real agentic benchmarks. A 1M token context window, coding benchmarks within 1 point of Claude Opus 5.5, and a deliberate, transparent approach to cyber safety controls. When the weights drop on November 4, it will be runnable on infrastructure users control, which matters for privacy, auditing, and sovereignty [11].

The cybersecurity disclosure is also notable. Most labs do not publish their internal safety evaluations in this level of detail. Kimi showing both the unfiltered and filtered capability numbers, explaining their reasoning, and committing to gradual diffusion of dual-use capability is the kind of transparency that should be the industry standard [12].