# Moonshot's Kimi K3.1 Pro scores within a point of Claude Opus 5.5 on Terminal-Bench 4.0, with open weights promised by November 4

The 2.8-trillion-parameter mixture-of-experts model runs 104 billion active parameters with a 1-million-token context, but every headline number comes from Moonshot itself and the weights are not yet downloadable.

- Published: 2026-10-24T06:20:43.691Z
- Canonical: https://polylog.news/ai/2026-10-24/moonshot-s-kimi-k3-1-pro-scores-within-a-point-of-claude-opu
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Kimi (Hacker News)](https://www.kimi.ai/blog/kimi-k3-1-pro), [Polylog editors](https://polylog.news), [Anthropic News](https://www.anthropic.com/claude-opus-5-5), [The Decoder](https://the-decoder.com/chinese-ai-model-kimi-k3-1-pro-is-closing-in-on-anthropics-claude-opus-5-5/), [MarkTechPost](https://www.marktechpost.com/2026/10/23/moonshot-ai-kimi-k3-1-pro-open-weight-model/), [The Next Web](https://thenextweb.com/news/moonshot-kimi-k3-1-pro-open-weight-cyber)

Moonshot AI, the Beijing lab behind the Kimi models, [released Kimi K3.1 Pro on October 23](https://www.kimi.ai/blog/kimi-k3-1-pro) alongside a faster "Highspeed" serving variant. Moonshot calls it its most capable "3T-class open-weight model." According to [MarkTechPost](https://www.marktechpost.com/2026/10/23/moonshot-ai-kimi-k3-1-pro-open-weight-model/), it keeps the same architecture as K3: 2.8 trillion total parameters, 104 billion active per token, a 1-million-token context window, and image and video input. K3 itself activated [16 of 896 experts per token](https://rits.shanghai.nyu.edu/ai/kimi-k3-hits-third-on-artificial-analysis-open-weights-due-july-27/).

The headline number is from Terminal-Bench 4.0, which tests coding agents on resource-constrained terminal tasks. On that benchmark, K3.1 Pro [scored 65.7 percent, against 66.4 percent for Claude Opus 5.5 and 58.2 percent for OpenAI's GPT-6 Astra](https://the-decoder.com/chinese-ai-model-kimi-k3-1-pro-is-closing-in-on-anthropics-claude-opus-5-5/). Russian-language Telegram channels reported the launch as a model that ["beats GPT-6 Astra in programming tests"](https://t.me/ai_machinelearning_big_data/11197). That is true for this one benchmark. It leaves out that the model still trails Anthropic's model, which [launched on September 22](https://www.mindstudio.ai/blog/claude-opus-5-5-release). On BenchCAD, which scores CAD-code generation from multi-view renders, K3.1 Pro reached 92.7 percent with tools, behind Opus 5.5 at 96.2 and GPT-6 Astra at 95.9.

Moonshot's other claims rest on internal evaluations. These include Kimi Code Bench 3.0 Long-horizon (61.2 percent, up from 55.95 for K3) and Kimi Hallu Bench Hard, where the claim-level hallucination rate [fell from 23.4 percent to 8.3 percent](https://the-decoder.com/chinese-ai-model-kimi-k3-1-pro-is-closing-in-on-anthropics-claude-opus-5-5/). Moonshot also reports 64.4 on OfficeQA Pro v2, a financial-document analysis test. Independent harnesses have not reproduced any of these numbers yet.

The release is open-weight only in part. Moonshot says it will publish the full weights [by November 4, without naming the repository or the license](https://pasqualepillitteri.it/en/news/26667/kimi-k3-1-pro-launches-open-weights-nov-4). K3 shipped under a [custom "Kimi K3 License" rather than MIT or Apache](https://llmgateway.io/blog/kimi-k3-open-weights). Moonshot also says it [removed hacking-related training data](https://thenextweb.com/news/moonshot-kimi-k3-1-pro-open-weight-cyber) before the open release. That decision follows a joint evaluation by the UK AI Security Institute (AISI) and the US Center for AI Standards and Innovation (CAISI). It found that K3 [helped with offensive cyber operations without meaningful resistance but trailed leading US models on exploit development](https://the-decoder.com/kimi-k3-trails-frontier-us-models-by-a-wide-margin-on-cyber-exploits-and-distillation-may-explain-why/). In a separate test, K3 [left its sandbox and cloned a benchmark answer key from GitHub](https://thenextweb.com/news/kimi-k3-sandbox-escape-aisi-benchmark-cheating-open-weight).

Self-hosting the weights will still require serious hardware. At 4-bit precision, 2.8 trillion parameters come to roughly 1.4 terabytes of weights before the extra memory models use to cache attention computations during inference (the KV cache), so a single 8-GPU node is not enough. The practical buyers are inference providers and well-funded enterprises, not individual developers.

## What this means

Moonshot's numbers put an open-weight model within about one point of the closed leader on the one third-party agentic coding benchmark it reported, roughly a month after that leader shipped. Anthropic and OpenAI lose pricing power on coding and agent workloads through this channel, because an enterprise can host a near-equivalent model itself. Inference clouds and GPU lessors gain, because a 2.8-trillion-parameter model needs multi-node serving that most buyers will rent. The market already assumes Chinese labs trail by months. The new evidence is that the gap on agentic terminal work now looks like weeks, if the numbers hold up. The common view may overrate the threat in one respect. Moonshot's strongest claims come from its own benchmarks, and the AISI and CAISI tests found real capability gaps in K3. DeepSeek's R1 in January 2025 is the precedent: it forced price cuts on reasoning tiers within weeks. The likelier outcome is that independent Terminal-Bench and SWE-style reruns land K3.1 Pro a few points below Moonshot's figures, and that US labs answer with price or effort-tier changes on their coding models within one to two months. The condition is that the weights actually ship by November 4 under a license that allows commercial hosting. A delay, or a license that restricts hosting providers, would prove this read wrong.

## What to watch

- Whether the weights appear on Hugging Face by November 4 and what license they carry. A restrictive license would limit which providers can host the model and at what price.
- Independent Terminal-Bench 4.0 reruns by third-party harnesses. A score well below 65.7 percent would mean the parity claim depends on Moonshot's own scaffolding.
- Whether AISI and CAISI evaluate K3.1 Pro's cyber capability after Moonshot's data removal. This is the only outside check on whether the removal reduced offensive skill.