Kimi K3.1 Pro Review: 2.8T Open Weights, Near-Flagship Terminal Coding, Lagging Science Tasks

The architecture sets the bar for running it: 2.8 trillion total parameters does not mean every computation uses all of them. Per token, the router picks 16 of 896 experts plus 2 shared experts, for 104 billion active parameters; the text decoder has 93 layers, with attention combining KDA and Gated DeepSeek Sparse Attention, whose indexer selects up to 2,048 earlier positions per query; images pass through a 401-million-parameter MoonViT-V2 vision encoder; weights are MXFP4 and activations MXFP8, trained with quantization awareness.

1. Publisher and dates: Moonshot AI; released Oct 23, 2026; full weights published on Hugging Face Nov 4, 2026
2. Model: Kimi K3.1 Pro, the first in the K3.1 series, a post-trained version of K3; a natively multimodal agentic model; text and image input, text output
3. Scale and context: 2.8 trillion total parameters, 104 billion active per token, 1,048,576-token context window
4. Reasoning effort: the model card notes that all vendor-reported scores were obtained at the maximum reasoning effort, temperature 1.0 and top-p 0.95
5. Channels: Hugging Face repository moonshotai/Kimi-K3.1-Pro, 112 files as of Nov 9; API via the Kimi API (kimi-k3.1-pro and kimi-k3.1-pro-highspeed) and 18 providers; self-hosting via vLLM, SGLang and TokenSpeed as recommended in the model card
6. Price: API at $2.00 per million input tokens, $0.20 cached input and $10.00 output; the Highspeed tier at $4.00 input and $20.00 output, with 2 to 3 times the speed according to the vendor; 33% below Kimi K3 at $3.00 input and $15.00 output
7. Licence: Kimi K3.1 Pro Licence; operators whose model-as-a-service business exceeds $20 million in combined annual revenue need a separate agreement with Moonshot AI before commercial use; products with over 100 million monthly active users or over $20 million in monthly revenue must display “Kimi K3.1-Pro”

Coding is the strongest part of the vendor's comparison table: Terminal-Bench 4.0 reads 65.7%, above GPT-6 Astra at 58.2% and below Claude Opus 5.5 at 66.4%, a gap of 0.7 percentage points; SWE Marathon v1.1 reads 51.7%, above Opus 5.5 at 46.3% and GPT-6 Astra at 42.5%; DeepSWE v1.1 reads 75.5%, above Opus 5.5 at 74.2% and GPT-6 Astra at 74.1%, below DeepSeek V4.1 Flash at 77.0%; BenchCAD with tools reads 92.7%, below Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%, and without tools 72.8%, close to Opus 5.5 at 73.0%; the internal long-horizon Kimi Code Bench 3.0 reads 61.2%, above K3 at 55.95%.

Agentic and knowledge work form the other set of vendor figures: Humanity's Last Exam with tools (HLE-Full w/ tools) reads 70.0%, above Opus 5.5 at 67.7% and GPT-6 Astra at 57.2%; BrowseComp browsing retrieval reads 92.1%, above Opus 5.5 at 90.8% and GPT-6 Astra at 91.5%; Agents' Last Exam reads 43.2%, above Opus 5.5 at 42.6% and GPT-6 Astra at 41.8%; AutomationBench-AA reads 73.3%, above Opus 5.5 at 69.5% and GPT-6 Astra at 68.5%; GDPval-AA v2.1 reads 1,846 Elo, above GPT-6 Astra at 1,573 Elo and below Opus 5.5 at 1,866 Elo; OfficeQA Pro v2 financial-document analysis reads 64.4%, above Opus 5.5 at 62.2% and GPT-6 Astra at 54.4%, the highest in the vendor's table for that test; GPQA-Diamond reads 94.2%, above Opus 5.5 at 93.9% and below GPT-6 Astra at 96.1%. Factual accuracy is an internal-benchmark figure: Kimi Hallu Bench Hard with tools gives a claim-level hallucination rate of 8.3%, below K3 at 23.4% but above GPT-6 Astra at 5.5%.

The weak spots are stated in the same vendor table: Terminal-Bench Science 0.1 scientific terminal tasks read 38.1%, below Opus 5.5 at 63.3% and GPT-6 Astra at 68.1%; HLE-Diamond reads 39.4%, below Opus 5.5 at 60.9% and GPT-6 Astra at 65.8%; Program Bench reads 31.5%, below Opus 5.5 at 65.0% and GPT-6 Astra at 50.0%; MMMU-Pro reads 84.1%, below Opus 5.5 at 87.7%. The cybersecurity rows reflect a deliberate vendor downgrade: the internal xdev-bench-hard reads 81.0%, below the 90.5% of the checkpoint built without cybersecurity data removal; ExploitBench reads 74.7%, below Opus 5.5 at 91.0% and GPT-6 Astra at 100%. The vendor's tech blog states that the open-weight release removed cybersecurity-related data from mid-training and reinforcement learning while keeping general coding tasks.

Independent boards are read separately from vendor figures: the Artificial Analysis Intelligence Index v4.3.2 (article published Oct 23, 2026) reads 55, fourth overall, 11 points above K3 at 44, behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56); after the index moved to v5.0.1, the model page (checked Nov 9, 2026) reads 46, above the median of 23 for open-weight models of similar size. On Artificial Analysis's own harness, Terminal-Bench 4.0 reads 67%, above K3 at 13% and Claude Sonnet 5.5 at 64%, the highest score the evaluator had recorded at the time; GDPval-AA reads 1,846 Elo, third overall. LiveBench 2.0 (checked Nov 9, 2026) gives 71.3 overall, below Claude Fable 5.1 at 85.7 and GPT-6 Astra at 85.6; by category, reasoning 91.1%, coding 81.8%, browser use 75.9%, research 77.9% and long-horizon agentic work 39.1%; cost per successful task is $1.044, below Claude Opus 5.5 at $4.820 and GPT-6 Astra at $4.828. Speed is the other side of the independent record: Artificial Analysis measured 189.63 seconds to the first answer token and 37.2 minutes per Intelligence Index task, among the slowest of open-weight models of similar size.

The conversion is stated openly: one representative vendor-reported score per dimension, divided by 10 into a 10-point scale; the total is the mean of the five dimension scores. The converted scores only restate benchmark figures, with no editorial judgement added; independent board scores are listed separately and excluded from the total.

1. Agentic coding 6.6: Terminal-Bench 4.0 65.7 converted, from Moonshot AI's published model card figure
2. Long-horizon coding 5.2: SWE Marathon 51.7 converted, from Moonshot AI's published model card figure
3. Agentic overall 4.3: Agents' Last Exam 43.2 converted, from Moonshot AI's published model card figure
4. Office documents 6.4: OfficeQA Pro v2 64.4 converted, from Moonshot AI's published model card figure
5. Browsing retrieval 9.2: BrowseComp 92.1 converted, from Moonshot AI's published model card figure
6. Total 6.3: mean of the five converted dimension scores (6.6+5.2+4.3+6.4+9.2)/5
7. Independent boards, listed separately: Artificial Analysis Intelligence Index v5.0.1 46, and 55 on v4.3.2 at launch, fourth overall; Artificial Analysis's own Terminal-Bench 4.0 run 67%, the highest the evaluator had recorded at the time; LiveBench 2.0 overall 71.3, with reasoning 91.1%, coding 81.8% and long-horizon agentic work 39.1%; none counts toward the total

On function and price, the following teams can already obtain the weights through public channels or connect via the API: teams that need to run a 2.8T-class open-weight model on their own hardware for long-horizon coding and office-document agent tasks, with the model card recommending the vLLM, SGLang and TokenSpeed inference engines; teams using the API on a pay-per-use basis at $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.20 per million tokens; and teams that need the 1M-token context window for long documents and codebases.

In the other direction, teams needing scientific terminal tasks and hard-knowledge question answering face Terminal-Bench Science 0.1 at 38.1% and HLE-Diamond at 39.4%, each more than 20 percentage points below the two closed flagships, so re-testing on the target task comes before deployment; interactive uses sensitive to first-token latency must account for the 189.63 seconds Artificial Analysis measured to the first answer token; and operators offering access to the model as a model-as-a-service with combined annual revenue above $20 million need a separate agreement with Moonshot AI before commercial use.