Kimi K3.1 Pro model release illustration with terminal and code visuals

Moonshot Kimi K3.1 Pro Nears Claude Opus 5.5 on Terminal-Bench, Open Weights Due November 4

Moonshot AI released Kimi K3.1 Pro on October 23, a 3T-class open-weight model with a 1-million-token context that scores 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 and ahead of GPT-6 Astra, with full weights promised by November 4.

Model releasesMoonshotCoding AgentsBenchmarks

Moonshot AI released Kimi K3.1 Pro on October 23, the first model in its new K3.1 series and what the company calls its most capable 3T-class open-weight model to date. It is live now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with the full model weights promised by November 4.

The headline number is on Terminal-Bench 4.0, a benchmark that grades coding agents on realistic, resource-constrained terminal tasks with strict task-specific tests. K3.1 Pro scores 65.7 percent, ahead of OpenAI GPT-6 Astra at 58.2 percent and narrowly behind Anthropic Claude Opus 5.5 at 66.4 percent. All of these figures are vendor-reported, and Moonshot says K3.1 Pro was evaluated with the Claude Code harness at maximum reasoning effort.

What the model is

K3.1 Pro is a post-trained version of Kimi K3, with substantially larger post-training compute. Moonshot describes it as a 3T-class mixture-of-experts model with a 1-million-token context window, accepting text, image and video input, with reasoning always on. Third-party write-ups put the size at 2.8 trillion total parameters with about 104 billion active per token.

A faster variant, Kimi K3.1 Pro Highspeed, ships alongside it at 2 to 3 times the speed for more responsive interactions. API pricing is 2 dollars per million input tokens and 10 dollars per million output tokens, according to Moonshot pricing pages cited in launch coverage.

Beyond coding, Moonshot reports gains across its internal benchmarks. On BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, K3.1 Pro scores 92.7 percent with tools, approaching Claude Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent. On OfficeQA Pro v2, a test of complex fiscal data analysis over real financial documents, it scores 64.4, which Moonshot calls state of the art. On its long-horizon coding benchmark, Kimi Code Bench 3.0, it scores 61.2 percent, up from 55.95 percent for K3.

The company also reports a sharp drop in hallucinations on its own stress test: the claim-level hallucination rate fell from 23.4 percent for K3 to 8.3 percent for K3.1 Pro. Moonshot cautions that this is a deliberately adversarial stress test, not a measure of everyday error rates.

The cybersecurity section is the interesting part

Unusually for a model launch post, a large section of Moonshot’s blog is devoted to cybersecurity. As K3.1 Pro grew more capable, Moonshot says it traced rising vulnerability-research performance to cyber-related content embedded in its general mid-training corpus and cyber tasks in its reinforcement-learning environment mix. For the open-weight release, it identified and removed cyber-related data from both training stages, while keeping the general-purpose coding tasks needed for usefulness.

The effect is measurable on Moonshot internal xdev-bench-hard, a set of expert-curated vulnerability discovery and exploitation tasks that reproduce recently disclosed flaws in software such as the Linux kernel, Node.js and JavaScriptCore. Without the data removal, the same pipeline solves 90.5 percent of tasks; the final K3.1 Pro solves 81.0 percent. That is still a large jump from K3 at 33.3 percent, which Moonshot attributes to stronger general coding and debugging ability generalizing to cyber tasks even after targeted data removal.

Moonshot states two principles behind the decision. First, it aims not to release open weights that extend the absolute frontier of cyber capability: K3.1 Pro offensive cyber capability is, by design, below what its pipeline can produce and behind frontier proprietary models. Second, it intends to progressively raise the cyber capability available in open weights, arguing that letting the capability diffuse gradually, rather than concentrating it in a few proprietary systems, ultimately makes all software more secure.

Why it matters

If the weights land by November 4 as promised, K3.1 Pro would become the highest-scoring open-weights model on the Artificial Analysis intelligence index, shrinking the gap between open and proprietary models from 12 points to about 3, according to The Decoder analysis of the index. The license has not yet been published, and K3 weights run to roughly 1.6 terabytes, so this is a model for GPU servers, not desktops.

The timing also matters for the open-weight race. DeepSeek, Qwen, GLM and now Moonshot have all pushed large open models in 2026, and each release narrows the practical gap between what anyone can download and what the best closed labs sell. K3.1 Pro is the first open-weight model to sit within a point of the best proprietary model on a major agentic coding benchmark.

Two caveats are worth keeping. First, every benchmark number here is Moonshot’s own, run at maximum reasoning effort; independent results have not yet arrived. Second, the open-weight promise is a promise: until November 4, K3.1 Pro is an API product, not an open model. The release to watch is the weights.

It is also worth remembering what K3.1 Pro replaces. Kimi K3, released earlier in 2026, already pushed open models into serious agentic coding work, but it trailed the best proprietary models by a wide margin on the hardest benchmarks. K3.1 Pro closes most of that gap in a single post-training step, which says as much about how fast reinforcement learning at scale is improving as it does about Moonshot itself.