Moonshot has unveiled Kimi K3.1 Pro, presented as a multimodal, 3T‑class mixture‑of‑experts (MoE) model. According to the company, the model has 2.8 trillion total parameters, 104 billion active parameters per token, and supports a 1‑million‑token context window. Moonshot highlights capabilities such as full‑project coding, parametric work in Blender/CAD, financial modeling, and long‑running agent workflows.
At launch the API and product variants went live: Standard and Highspeed editions appeared on Kimi’s website and app, in Kimi Work and Kimi Code, and via the API. Moonshot claims the Highspeed tier preserves model quality while generating outputs 2–3× faster; the company’s summary says Highspeed is roughly twice the price. The release also introduced Kimi Hosted Agents, a managed runtime for 24/7 execution and multi‑agent collaboration.
What “open” means — weights and license
Moonshot’s announcement pledges to publish model weights on November 4, 2026. Public posts and tweets, however, do not specify final license terms, redistribution rules, commercial‑use restrictions, checkpoint precision, or full inference requirements. Because of that uncertainty, it is more accurate to call K3.1 Pro a ‘‘soon‑to‑be open‑weight’’ model rather than unqualified ‘‘open source.’’ Artificial Analysis explicitly noted the weights as unpublished at the time of its evaluation.
Independent evaluations and measured performance
Artificial Analysis (AA) placed K3.1 Pro at 55 on its Intelligence Index, ranking it fourth (Opus 5.5 was 58; Sonnet 5.5 and Muse Wonder Full were 56). AA reported that most of the model’s improvements were agentic (agent‑style tasks) rather than uniformly distributed across knowledge and science tests.
Arena placed K3.1 Pro at #8 in its WebDev ranking with 1,692 points — nine places and 37 points better than Kimi K3. On cyber metrics, K3.1 Pro scored 64% on AA’s Cyber Index, the highest publicly available score and second overall behind trusted‑access GPT‑6 Sol Daybreak Blue at 69%.
AA also reported that accuracy on AA‑Omniscience rose from 48% to 51% and hallucination rate fell from 53% to 44% when comparing K3 to K3.1 Pro — a nine percentage‑point reduction in that measurement. Moonshot has advertised larger reductions (for example a 65% figure for finance work) but did not publish the methodology for that claim.
Cost, token pricing and workload dependence
Moonshot’s list pricing for Standard first‑party access is: $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached‑input tokens — roughly one‑third lower than K3’s prior $3/$15/$0.30 pricing. The company’s Chinese account published RMB equivalents. Highspeed is described as costing about twice the standard price.
Independent task‑level cost estimates show workload dependence: AA reported an Intelligence Index median cost of about $1.98 per task (roughly unchanged from K3), while Arena reported a median WebDev cost of $1.23 per task (29% below K3). On AA’s cyber suite, K3.1 Pro had higher task costs driven by roughly double input‑token consumption.
Deployment and local‑model considerations
A trillion‑parameter MoE does not imply simple local deployment because the expert pool plus runtime state must be handled. As an illustrative calculation (not a measured configuration), 2.8T parameters could imply about 5.6 TB of storage at 16‑bit precision or around 1.4 TB at 4‑bit, before quantization metadata and serving overhead. The publicly shared posts disclose no layer layout, expert routing, attention design, training‑compute budget, or measured full‑context serving footprint — all of which matter for claims about ‘‘running locally.’’
Safety intervention and cyber training data removal
Moonshot reported removing cyber‑related data during mid‑training and reinforcement learning, rather than relying only on inference‑time refusals. An internal score cited in posts fell from 90.5% to 81% (a 9.5 percentage‑point reduction) while other capabilities improved. Early commentators called the model ‘‘nerfed’’ for cyber tasks; others argued the intervention is a deliberate trade‑off to limit downstream offensive capability.
AA observed that public defensive tasks were still answered without refusals. The difference between training‑time capability reduction and API refusal policy was widely noted: a lower internal capability (relative to a counterfactual) is not the same thing as the model refusing specific API requests. AA also warned that proprietary trusted‑access variants can have substantially different public vs. private defensive scores.
Demonstrations and task‑specific results
Moonshot provided demonstrations for end‑to‑end work in Unreal, Unity, Godot, Blender and CAD and claimed a 65% reduction in hallucinations for finance work; the vendor did not publish baseline definitions or evaluation protocols for that finance figure. Demo items included a 3D robotic hand decomposed into 2,317 pieces and a Blender model of Singapore Changi Airport with assets, scene assembly and rendering.
Independent evidence beyond conventional coding and agent tasks is mixed. AA’s preliminary 3D voting suggested K3.1 Pro may rank below its coding/knowledge standing in 3D tasks. Other third‑party and user anecdotes reported comparable or mixed results versus peers for game‑related and spatial tasks. These are workload‑dependent observations rather than universal conclusions.
Benchmark interpretation and comparative caveats
Several observers stressed that Moonshot’s benchmark table is not an independent replication. Harness dependence (choice of agent harness and tools), inconsistent attribution of comparison data, benchmark versioning and settings can all materially affect rankings and cost‑performance claims. Some commentators argued the launch narrowed the accessible frontier gap; others emphasized that competitiveness remains task‑dependent and questioned inference economics.
Promotion allegations and launch credibility
Late in the coverage, allegations appeared about instructions for undisclosed paid promotion. One commentator claimed he did not receive or accept payment and called for an apology; the alleged briefing text circulated in posts. No Moonshot response or independent authentication of the alleged briefing was available in the collected material, so these remain unverified transparency allegations separate from independent benchmarking results.
What this means for the field
Kimi K3.1 Pro represents a notable incremental advance for Moonshot’s Kimi line: larger post‑training compute, measurable agentic gains, and competitive independent scores in some indices. The practical impact depends on the November 4, 2026 weight release, the eventual license and redistribution terms, community reproduction of benchmarks, and the real‑world serving economics of a 2.8T MoE model. Community reaction mixed technical recognition with substantive concerns about deployment costs, benchmark transparency and promotional practices.
Key numbers and dates
- Total parameters: 2.8 trillion
- Active parameters: 104 billion
- Context window: 1,000,000 tokens
- Artificial Analysis Intelligence Index: 55 (4th place)
- Arena WebDev: #8, 1,692 points
- AA Cyber Index: 64%
- Promised weights publication: November 4, 2026
- Standard list pricing: $2/M input, $10/M output, $0.20/M cached‑input
- Coverage and community discussion concentrated around October 22–23, 2026
Further assessment will require the actual weight release, transparent license terms, and independent reproductions of the reported benchmarks and cost measurements.
