The IT Home article on Kimi K3.1 Pro prints the visible time 2026/11/5 8:00:15. Read as Beijing time, that is 08:00:15 on Nov. 5, one second after the start of this noon window. The byline and the editor credit are both Wen Zhou. The article says Moonshot, following a promise made at the Oct. 24 launch, opened the Kimi K3.1 Pro weights on Hugging Face today. The repository record is not today. For moonshotai/Kimi-K3.1-Pro, createdAt and lastModified are both 2026-11-02T04:13:51.000Z, which is 12:13 in Beijing on Nov. 2, earlier than this edition's backfill line. "Today" is the article's word. It is not the clock on the hub record.
[1][2]
The article then cites Agent Arena, operated by Arena. It ranks Kimi K3.1 Pro (Max) first among open-weight models and 13th among all models entered, and first among open models in work, coding, and chat. The ranking is said to rest on more than 17,500 real long-horizon agent sessions worldwide, with a net improvement rate of +5.1 percent, against +0.9 percent for the previous K3 (Max), a gap of 4.2 percentage points. Bash Reliability is said to move from -4.5 percent for K3, ranked 13th, to +7.9 percent, first among open models. Praise versus Complaint moves from +1.3 percent to +7.7 percent. Steerability is +6.5 percent, also first among open models. Confirmed Success is +3.9 percent, fourth among open models. The article describes the measurement as signals taken from real sessions, including confirmations, praise, complaints, corrections, and command errors, then a randomized inference that Arena calls causal tracing. This column does not expand that procedure. The article does not say on which day the leaderboard snapshot was taken.
[1][2]The model is described as a mixture of experts, with 2.8 trillion parameters in total, about 104 billion active on a pass, and a context of about 1 million tokens. The article separates "can be downloaded" from "can be run locally": the BF16 weights are about 5.6 TB, INT4 still about 1.4 TB, and a single consumer graphics card basically cannot run them. On Terminal-Bench 4.0, the article gives K3.1 Pro 65.7 percent, GPT-6 Astra 58.2 percent, and Claude Opus 5.5 66.4 percent, a gap of less than one percentage point. That sentence is in the prose. The large table that follows does not date the tests, and a related link points back to an earlier launch story. This column does not recast the rest of that table as a new measurement from today, and it does not lift the rows that concern exploit benchmarks.
[1][2]The article lists four kinds of signal: an explicit confirmation that a task is done, praise and complaints, correction instructions, and errors from shell commands. The net improvement rate comes after those. The four kinds are the leaderboard's own signals, not a recount by IT Home. The article places the item under a Nov. 5 dispatch and points back to the Oct. 24 promise. The promise date is in that sentence. The repository record's date is not. Both dates stay.
[1][2]The article adds two sentences about the model itself, with no independent replication attached. Against K3, K3.1 Pro is described as a large joint reinforcement-learning post-training run on the K3 base, covering hundreds of downstream fields, with some full-attention layers replaced by gated sparse attention in order to lower the cost of long context. The sentence does not say how many layers changed or what data the post-training used. Agent Arena is described as different from a paired blind test: it does not ask a person to vote between two answers, and the signals come from real sessions. That is the article's summary of the leaderboard method. This column does not fill in how the randomized inference samples.
[1][2]What can be fixed is the account IT Home published at 08:00:15, and a hub record that stops on Nov. 2. The rank, the net improvement rate, and the Terminal-Bench score are the article's figures, not an independent replication. The 17,500 sessions are given no date range. The license text and the exact way active parameters are counted are not spelled out far enough to check. The social thumbnail shows the repository name, not this leaderboard.
[1][2]