# Kimi K3.1 Pro Kimi K3.1 Pro is the first model of [[Moonshot AI]]'s K3.1 series, announced around 23 October 2026. It's [[Kimi K3]] with a lot more post-training compute, not a new architecture: same 3T-class mixture of experts (Artificial Analysis lists 2.8T parameters, an HN commenter cites 104B active), same 1M token context. Moonshot calls it "our most capable 3T-class open-weight model to date". It comes in two flavors: - **Kimi K3.1 Pro** (`kimi-k3.1-pro`) - **Kimi K3.1 Pro Highspeed** (`kimi-k3.1-pro-highspeed`): same model, 2-3x faster according to Moonshot. No tokens-per-second figure or separate price in the announcement ## Where you can use it - kimi.com and the Kimi apps (iOS, Android, HarmonyOS) - Kimi Work desktop app (3.1.0 or later, Windows and Apple silicon Macs) - [[Kimi Code]], with `/model` - The [[Kimi|Kimi API]], with Zero Data Retention available for API customers - Kimi Enterprise, off by default at launch: admins have to turn it on - Open weights: "The full model weights will be released by Nov 4." No license mentioned yet ## Pricing $2 per million input tokens and $10 per million output tokens, $0.20 for cached input (Artificial Analysis lists a 90% cache discount). That's cheaper than K3 ($3 / $15 / $0.30) for a model that's the same size or slightly bigger, and half the price of [[Claude Opus 5.5]] ($4 / $20). The subscription side is a different story. The most common complaint in the HN thread is the 5-hour usage window on Kimi's plans. Several people on the $99-$100 per month plan said they hit the limit in under an hour, one "with one prompt". Others said the quota had been cut with each release since K2.7. These are user reports, not official numbers, but there are a lot of them. Commenters who mostly care about cost per task still prefer [[DeepSeek V4.1 Flash]], which is cheaper and faster, and "good enough 95% of the time" according to one of them. ## Benchmarks Moonshot's own numbers, at max reasoning effort: | Benchmark | K3.1 Pro | Comparison | |---|---|---| | Terminal-Bench 4.0 | 65.7% | [[Claude Opus 5.5]] 66.4%, [[GPT-6 Astra]] 58.2% | | Kimi Code Bench 3.0 long-horizon (internal) | 61.2% | K3 55.95% | | BenchCAD with tools (voxel IoU) | 92.7 | Opus 5.5 96.2, Astra 95.9, [[Claude Fable 5.1]] 92.6 | | OfficeQA Pro v2 | 64.4 | Moonshot calls it state of the art | | Kimi Hallu Bench Hard (claim-level hallucination rate, lower is better) | 8.3% | K3 23.4%, Astra 5.5%, Opus 5.5 15.0% | | Internal business operation | 37 | K3 8.9 | The hallucination drop (23.4% to 8.3%) and the business operation jump (8.9 to 37) are the biggest changes over K3. Both are internal benchmarks, so treat them as direction, not proof. Independent numbers: the HN submission and several comments cite a score of 55 on the [[Artificial Analysis]] Intelligence Index, 4th overall at the time and close to Opus 5.5, with the model sitting on the price/intelligence Pareto line. The same comments say it used 150M output tokens to run the index (median 87M), so it's chatty. When I checked the Artificial Analysis model page on 3 November, it showed 46 (23rd of 60) and a slow 31.6 tokens per second. I don't know whether that's a different reasoning setting or a revision. Commenters pointing at AA-Omniscience (the knowledge and hallucination index) put K3.1 roughly six months behind the frontier. ## Cyber capability, on purpose lower This part is new. Moonshot removed cyber-related data from mid-training and RL for the open-weight release, and says it aims "not to release open weights that would extend the absolute frontier of cyber capability". On its xdev-bench-hard, K3.1 Pro scores 81.0, against 90.5 for an internal Kimi model without that risk reduction and 33.3 for K3. Moonshot says the result "remains behind that of frontier proprietary models that have been publicly available for at least three months". It's the first time I see a Chinese open-weight lab publish a deliberate capability cut like this. ## The distillation debate, again The model sometimes calls itself ChatGPT in its reasoning, and one commenter measured its UI designs as more similar to OpenAI's than K3's were (0.58 vs 0.43). Some read that as [[Knowledge Distillation|distillation]] from US models. Others pointed out that public chat transcripts in the training data produce the same effect, and that ChatGPT and Claude have also called themselves DeepSeek. The [[Kimi K3]] architecture work (KDA, NoPE, LatentMoE) still makes "only distilled" a hard claim to defend. "Distilled in part, through post-training data" is the more likely reading, and nobody in the thread had proof either way. ## My take The API price is the story: frontier-adjacent quality, a 1M context, and open weights coming, for half of Opus 5.5. Through [[OpenRouter]] or another provider, it's a serious option for long agentic runs where you pay per token. I'd stay away from the Kimi subscriptions for now. A plan that runs out in 47 minutes isn't a coding plan, and K3.1 Pro spends a lot of reasoning tokens, which makes the limits hit even faster. The weights are a different matter. Like K3, running it yourself means more than a terabyte of memory, so "open" mostly means other providers can host it (and compete on price), plus [[AI Expert Offloading]] experiments for people with big machines. ## References - Moonshot AI, "Kimi K3.1 Pro" tech blog: https://www.kimi.ai/blog/kimi-k3-1-pro - Artificial Analysis model page: https://artificialanalysis.ai/models/kimi-k3-1-pro - Hacker News, launch discussion: https://news.ycombinator.com/item?id=50212065 - Hacker News, "Kimi K3.1 scores 55 on Artificial Analysis, placed on the Pareto line": https://news.ycombinator.com/item?id=50213337 ## Related - [[Kimi K3]] - [[Moonshot AI]] - [[Kimi]] - [[Kimi Code]] - [[Artificial Analysis]] - [[Claude Opus 5.5]] - [[GPT-6 Astra]] - [[Claude Fable 5.1]] - [[DeepSeek V4.1 Flash]] - [[Knowledge Distillation]] - [[AI Open Weight Models]] - [[AI Expert Offloading]] - [[AI Mixture of Experts (MoE)]] - [[OpenRouter]]