News · 2026-10-24

Kimi K3.1 Pro reaches fourth on an independent index, with weights promised for November 4

Moonshot launched Kimi K3.1 Pro on October 23, and Artificial Analysis placed it fourth on its independent Intelligence Index in its launch assessment. The model is available through hosted services now, with full weights promised by November 4. Its strong position makes it a serious alternative for coding and other knowledge work, but an independently scored service is still different from a downloadable, licensed checkpoint.

Key facts

• Artificial Analysis placed K3.1 Pro fourth in its October 23 launch assessment.
• Moonshot lists standard access at $2 per million input tokens and $10 per million output tokens.
• The model retains a one-million-token context window; weights are promised by November 4, 2026.
• Primary source: Moonshot’s launch announcement.

The release’s central change is more training after the initial model was built. Moonshot describes K3.1 Pro as Kimi K3 with substantially more post-training computation, including reinforcement learning across hundreds of domains. It also describes improvements to the systems that schedule training tasks, run sandboxes, and serve long conversations. The claim is that this combination makes additional training produce more predictable improvements. The announcement does not publish a full training recipe or total compute budget.

A useful analogy is an already capable apprentice receiving many more carefully checked assignments. The underlying foundation remains K3, while the practice, feedback, and workshop improve. That analogy explains the company’s account; it does not prove which individual change caused which gain. The site’s lesson on reward-based post-training explains why that distinction matters.

The strongest outside evidence is Artificial Analysis’ launch assessment. Its fourth-place result puts K3.1 Pro behind Claude Opus 5.5, Claude Sonnet 5.5, and Meta’s Muse Wonder Full in that dated assessment. Broad index proximity is meaningful, but it does not establish equal performance on every job. Moonshot’s terminal-coding comparison uses a particular agent setup and comparison scores drawn from different published sources; it is not a neutral fresh rerun of every model under identical conditions.

Price has a similar qualification. The official platform lists lower standard token rates than K3. Artificial Analysis nevertheless found roughly similar task costs because K3.1 Pro produced more output. Paying less per word does not guarantee paying less per finished report. Moonshot also offers a more expensive Highspeed version, described as retaining the same intelligence and context while running roughly two to three times faster. Buyers should compare completed work, waiting time, and repeated context reading together.

The release also makes a large factuality claim. On Moonshot’s internal hard test, the fraction of factual claims classified as hallucinated fell from 23.4% to 8.3%. Moonshot explicitly distinguishes this stress test from ordinary usage. Artificial Analysis’ separate factuality evaluation supports improvement, but on a different measure and with a smaller movement. Neither result supports a general statement that only 8.3% of normal Kimi answers contain errors. Our guide to AI benchmarks explains how changing the denominator changes the story.

Outside practical reception is mixed. In a six-assignment video test, AICodeKing placed the model around the middle of his tested field, reporting both useful successes and unfinished products. That is valuable dissent from one creator’s workflow, with selected tasks and a particular agent harness. It does not invalidate the broad index, just as the index cannot settle his workflow. The Hacker News discussion includes similar arguments about task selection, subscription access, and the gap between a leaderboard and sustained autonomous work.

The dated commitment is unusually clear: Moonshot says, “Full model weights will be released by Nov 4, 2026.” The company’s model listing had no K3.1 Pro checkpoint in the dossier’s October 24 check. Its disk download size and required graphics memory are therefore unpublished; parameter counts cannot supply either answer. Moonshot has also not specified the license, so calling today’s release open source would outrun the evidence.

There is one more reproducibility issue to watch. Moonshot says its open-weight release will remove cyber-related data while retaining general coding capability. The eventual downloadable model may therefore differ from the hosted system that earned the launch scores. For developers, that makes the November release a new evaluation target rather than a file-export milestone. The public variant will need its own practical checks on coding quality, factual reliability, and task cost. The practical next test is whether the shipped weights, license, hardware guidance, and independently reproduced results deliver the access implied by the announcement.

Primary source, verified: read the paper →

Ground Truth. (2026, October 24). Kimi K3.1 Pro reaches fourth on an independent index, with weights promised for November 4. Ground Truth. https://groundtruth.day/news/kimi-k3-1-pro-frontier-score-weights-november-promise.html

@misc{groundtruth:kimi-k3-1-pro-frontier-score-weights-november-promise,
  title  = {Kimi K3.1 Pro reaches fourth on an independent index, with weights promised for November 4},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/kimi-k3-1-pro-frontier-score-weights-november-promise.html}
}

Topics: models · moonshot · coding-agents · post-training · open-weights