In short: On 23 October 2026, Moonshot AI launched Kimi K3.1 Pro, the most capable model in its Kimi K3 series and one of the largest open-weight releases to date: a mixture-of-experts system with roughly 2.8 trillion parameters (104 billion active per token) and a one-million-token context window, according to the company's official announcement on LinkedIn. The model is already live on Kimi.ai, in the Kimi apps and via the API, with full open weights promised by 4 November 2026. On the company's reported Terminal-Bench 4.0 scores, K3.1 Pro reaches 65.7% — just behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2% — although independent benchmark readings are more measured.
What Moonshot AI announced on 23 October
According to Moonshot's official announcement posted on LinkedIn on 23 October 2026, Kimi K3.1 Pro is its flagship release: a 3-trillion-parameter-class open-weight model built on a mixture-of-experts (MoE) architecture. SiliconANGLE's reporting the same day put the figures at 2.8 trillion total parameters, of which 104 billion are active for each token — the MoE design means only a fraction of the network fires on any given token, which keeps inference costs below what a dense model of that size would require.
The model ships with a one-million-token context window, and it is available immediately across three surfaces: the Kimi.ai website, the Kimi apps, and the developer API under the model names kimi-k3.1-pro and kimi-k3.1-pro-highspeed. A Highspeed variant promises two to three times faster inference for latency-sensitive workloads.
Pricing
API pricing, as published with the launch, is $2 per million input tokens and $10 per million output tokens, with cached-input pricing at $0.20 per million tokens. The Highspeed tier costs $4 per million input and $20 per million output tokens — roughly double the base rate for the two-to-three-times speed gain. As always with vendor-published pricing, treat these figures as current at launch; API catalogues are revised regularly.
The benchmarks: strong on agentic coding, with caveats
Terminal-Bench 4.0 scores (company-reported) — hover or tap each bar for details.
The headline benchmark is Terminal-Bench 4.0, an agentic terminal-coding benchmark. Per the company's reported figures: Claude Opus 5.5 at 66.4%, Kimi K3.1 Pro at 65.7%, and GPT-6 Astra at 58.2%. On paper, that puts the new open-weight model within striking distance of the leading closed model on this particular task family.
But benchmark claims from a launching lab deserve the usual discount. An independent reading comes from Artificial Analysis's Intelligence Index, which scores Kimi K3.1 Pro at 55 against 58 for Claude Opus 5.5 — a similar ordering, but a wider gap than the company's own Terminal-Bench numbers suggest. Moonshot also cites 92.7% on BenchCAD, a CAD-design benchmark, and claims a 65% reduction in hallucination rates on finance tasks versus the earlier Kimi K3. Both of those figures are company-reported and, at the time of writing, not yet reproduced by independent evaluators.
The fair reading: K3.1 Pro looks genuinely competitive on agentic and engineering workloads, close enough to the frontier that the comparison is interesting — but the exact deltas should be treated as reported, not settled.
Why this matters
Three things make this launch worth paying attention to. First, scale: a 2.8-trillion-parameter open-weight model is a statement that the frontier is no longer the exclusive property of closed labs — provided the weights actually arrive by the promised 4 November date. Second, economics: at $2/$10 per million tokens with $0.20 caching, Moonshot is pricing aggressively against Western frontier APIs, which puts direct pressure on the cost curve for heavy agentic workloads. Third, timing: a one-million-token context window plus a Highspeed tier suggests the model is aimed squarely at production coding agents and long-context enterprise tasks, not demos.
There is also a strategic read. Open-weight releases from Chinese labs have become a regular feature of the model calendar, and each one shifts the "build versus rent" calculation for developers worldwide. If the weights land on schedule and independent evaluations reproduce the reported scores, K3.1 Pro becomes a credible self-hosted alternative to closed frontier APIs for engineering-heavy workloads.
What to watch
The single most important date is 4 November 2026, when Moonshot has promised the full open weights. A weights release is verifiable in a way launch-day benchmarks are not: the community will immediately start reproducing the reported numbers, fine-tuning derivatives, and probing the model's limits. Watch for independent reproductions of the Terminal-Bench and BenchCAD figures in the days after release, and for the first signs of how the Highspeed tier's pricing holds up under real load.
Frequently Asked Questions
Is Kimi K3.1 Pro fully open source yet?
Not at launch. The model is available through Kimi.ai, the Kimi apps and the API, and Moonshot has promised the full open weights by 4 November 2026. Until the weights are published, the model can be used but not self-hosted.
What does "104B active per token" mean?
Kimi K3.1 Pro uses a mixture-of-experts design: the network has about 2.8 trillion parameters in total, but only a routed subset — around 104 billion — is activated for each token. This keeps the compute cost per token far below what a dense 2.8T-parameter model would require.
How does its pricing compare to frontier APIs?
At launch, Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.20. The Highspeed tier doubles those rates for two to three times faster inference. These are launch-day figures from the company and may change.
Sources: Moonshot AI's official announcement (LinkedIn); SiliconANGLE, 23 October 2026.
Be the first to share a useful question or thought.