Kimi K3.1 Pro arrives hosted as a harder index redraws the model gap
Kimi K3.1 Pro is available through hosted services, with weights promised by November 4; a revised independent benchmark puts its performance and the downloadable-model gap in a different light.
Moonshot’s Kimi K3.1 Pro is live in Kimi’s hosted products and API.
Kimi Code’s October 23 release notes confirm the rollout, with access and the million-token context window dependent on subscription tier. Moonshot says it will release the full weights by November 4. For an operator choosing between a service and a self-hosted model, that remains a future commitment. [1][2]
On October 27, Artificial Analysis published Intelligence Index v5. Kimi scored 46, compared with 58 for Claude Opus 5.5 at maximum effort. The highest-scoring models whose weights were actually available, DeepSeek V4.1 Flash 1017 and GLM-5.3 at maximum effort, scored 32. This distinguishes Kimi’s measured performance from the set of models a team could already download. [4]
Sources & details 9 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- Moonshot — Kimi K3.1 Pro
Hosted availability and promise to release full weights by November 4.
- Kimi Code — October 23 release notes
Dated rollout and subscription-tier restrictions.
- Artificial Analysis — Kimi K3.1 Pro launch benchmarks
October 23 evaluation on the previous index; not directly comparable with v5.
- Artificial Analysis — Intelligence Index v5
October 27 scores and methodology; distinguishes promised weights from released weights.
- Artificial Analysis — Kimi K3.1 Pro model record
Current v5 score 46 and unreleased-weight status; a separate AA record, not another evaluator.
- Artificial Analysis — Claude Opus 5.5 model record
Current v5 maximum/default-fallback score 58; same evaluator as the v5 article.
- Artificial Analysis — DeepSeek V4.1 Flash 1017 model record
Current v5 maximum-effort score 32 and open-weight status.
- Artificial Analysis — GLM-5.3 release record
Current v5 maximum-effort score 32 and open-weight status.
- Artificial Analysis — Intelligence Index methodology
Separate version history confirms changed tasks and weights; does not turn cross-version scores into a longitudinal experiment.
Images & figures
Scores share AA’s v5 methodology. Kimi’s weights were promised by November 4, not downloadable at this check.
Euangelion / Codex · Source · Usage terms · Download figure