--- type: analysis slug: kimi-k3-1-pro-is-already-useful-but-open-weights-remain-the-real-test title: "Kimi K3.1 Pro is already useful, but open weights remain the real test" description: "Moonshot’s new model is available as a hosted service, with lower listed API rates than K3 and promising independent results in specific tests. But teams seeking deployment control still have to wait for the weights and license—and should compare task cost, not just token price." published_at: 2026-10-27T00:00:00.000Z topics: ["trends"] source_urls: ["https://www.marktechpost.com/2026/10/23/moonshot-ai-kimi-k3-1-pro-open-weight-model","https://www.kimi.ai/blog/kimi-k3","https://www.kimi.com/en/blog/kimi-k3-1-pro","https://www.kimi.com/blog/kimi-k3-1-pro","https://kimi.ai/blog/kimi-k3"] author: "Andrey Reshetnikov" author_profile: https://newruntime.com/owner-profile.md --- # Kimi K3.1 Pro is already useful, but open weights remain the real test Moonshot’s new model is available as a hosted service, with lower listed API rates than K3 and promising independent results in specific tests. But teams seeking deployment control still have to wait for the weights and license—and should compare task cost, not just token price. ## Summary Moonshot has released Kimi K3.1 Pro, a post-trained Kimi K3 mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters, and a reported 1,048,576-token context window. It is available through Kimi’s products and API, with listed rates below K3’s; Moonshot says a Highspeed variant runs two to three times faster. The company reports coding benchmark scores, while independent evaluations have produced mixed workload-specific results. Open weights are promised for November 4, 2026, but are not reported as available now. ## What happened K3.1 Pro accepts text, image, and video through the API, according to the release evidence. Moonshot reports 65.7% on Terminal-Bench 4.0 and 51.7 on SWE-Marathon; those are company-reported results, not independently validated in the evidence pack. The dossier adds two narrower external data points: according to its d8 finding, Artificial Analysis scored K3.1 Pro 55 versus 44 for K3 on its Intelligence Index, at about $1.98 per tested task. According to d9, a Plotly workload test recorded 42 of 43 tasks completed, $13.47 in cost, and a median 2.5 minutes; its K3 comparison was 35 of 43 tasks, $5.01, and 43 seconds. These results describe their respective test settings, not universal model rankings or costs. Moonshot says the Pro model used substantially more reinforcement-learning compute and that cyber-related material was removed from training and RL environments ahead of the planned open-weight release. ## What the archive adds Earlier K3 coverage in the archive records Moonshot’s stated plan to release K3 weights by July 27, 2026, alongside company claims about architecture and scaling efficiency. That is useful context for tracking follow-through, but it does not prove K3.1 Pro will be delayed. The dossier’s independent evaluations add a more practical frame than the launch benchmarks: measured quality, time, and cost can move in different directions depending on the workload. Neither narrow test establishes general performance or resolves whether K3.1 Pro weights will arrive with useful licensing terms. ## Trend signal Frontier-model launches increasingly split into two products: hosted access now and a possible portable model later. Kimi’s immediate API and app availability lets users test the model without managing infrastructure, while the planned weights promise a different kind of control if they arrive under workable terms. Independent workload tests underline why a single benchmark or token rate is insufficient: a model can complete more tasks yet take longer or cost more in a particular setup. The relevant unit for adoption is successful work at acceptable latency and total cost. ## Why now According to dossier findings d1 and d2, Moonshot’s existing API and product channels enabled immediate access through the API and Kimi Code. Finding d3 reports listed K3.1 Pro rates about 35% below K3 on standard input and output, giving users a lower-priced hosted entry point. The primary evidence says K3.1 Pro used substantially more reinforcement-learning compute; this is a stated training choice, not proof of a particular cost advantage. The dossier’s workload-specific evaluations help explain why lower token rates alone do not determine whether the model is economical for a task. ## Connections to other events The strategic question is whether K3.1 Pro’s hosted availability becomes a durable deployment option or stays a vendor service. The API rates are listed at $2 per million input tokens and $10 per million output tokens, compared with $3 and $15 for K3, but Plotly’s reported test shows that a specific workload can cost more even while improving task completion. Our read is that open weights would matter most for teams that need customization, deployment control, or predictable access—not simply for those seeking the lowest per-token price. Until the release and license are public, those teams cannot assume portability. ## Why it matters Builders can test K3.1 Pro now, but should evaluate it against their actual work rather than infer value from parameter counts, context length, or headline pricing. Buyers should separately assess hosted-model economics and the future open-weight proposition. For a coding or agent workflow, log completion quality, retries, tool errors, latency, and total spend; these measures reveal whether improved success justifies slower or costlier runs. Treat Moonshot’s cyber safeguards as company claims until independently assessed, especially when planning deployments that could expose the weights to broader use. ## What happens next - **An independent public evaluation will compare K3.1 Pro with K3 on at least one named coding or agent workload and report task success alongside time or cost.** (6 months, 70%). Watch for: A reproducible report publishes its workload, methodology, and comparative outcome measures. - **Moonshot will publish K3.1 Pro weights or clarify their release status and licensing by its stated target date.** (3 months, 65%). Watch for: A public download or official update specifies availability and license terms by November 4, 2026. ## Limits and uncertainty The evidence does not establish real-world throughput across hardware or workloads, and the Artificial Analysis and Plotly results are limited to their respective evaluation settings. Moonshot’s Terminal-Bench and SWE-Marathon results remain company-reported here. The dossier does not settle the final license, the completeness of the planned weights, or whether the November 4, 2026 target will be met. Earlier K3 release timing is context, not evidence of a K3.1 delay. This take could be wrong if broad independent testing shows consistently strong task economics and the weights arrive promptly on permissive terms. ## Skill to build now Build a workload-level model comparison: run one real coding task against your current model and K3.1 Pro, repeat failures, and record quality, retries, tool errors, latency, and total token cost. This week’s exercise is to produce a one-page scorecard that makes the trade-off between success rate and task economics visible. ## Reader takeaway This week, choose one representative coding or agent task and run K3.1 Pro alongside your current model. Record successful completion, retries, tool errors, end-to-end time, and total cost—not just tokens or benchmark score. Keep the result as a hosted-service baseline; revisit self-hosting only after the K3.1 Pro weights and license are actually published. ## Open question Will the eventual K3.1 Pro weights and license let teams obtain meaningful deployment control, and will independent tests confirm that the stated cyber-risk reductions persist beyond Moonshot’s training process? ## Sources - [Moonshot AI releases Kimi K3.1 Pro, a 2.8T-total/104B-active MoE model with 1M context](https://www.marktechpost.com/2026/10/23/moonshot-ai-kimi-k3-1-pro-open-weight-model) - [kimi.ai](https://www.kimi.ai/blog/kimi-k3) - [kimi.com](https://www.kimi.com/en/blog/kimi-k3-1-pro) - [Moonshot says Kimi K3.1 Pro was trained with cyber-related data removed to cap offensive capability before open-weight release](https://www.kimi.com/blog/kimi-k3-1-pro) - [kimi.ai](https://kimi.ai/blog/kimi-k3) - [Kimi API Open Platform](https://platform.kimi.com/) - [Kimi K3: Open Frontier Intelligence](https://arxiv.org/abs/2607.24653) - [Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance; it will be the leading open-weights model once its weights are published](https://artificialanalysis.ai/articles/kimi-k3-1-pro-launch-benchmarks) - [Kimi K3.1 Pro Benchmark Results for Data Analysis](https://plotly.com/blog/kimi-k3-1-pro-plotly-data-analytics-bench/) - [Kimi K3.1 | OpenLM.ai](https://openlm.ai/kimi-k3.1/) - [Kimi K3.1 Pro](https://www.ridgebench.com/models/kimi-k3.1-pro) - [Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed are available now in AI Gateway](https://vercel.com/changelog/kimi-k3-1-pro-and-kimi-k3-1-pro-highspeed-are-available-now-in-ai-gateway) Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json ## Continue in the graph Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`. - topic hub: [Trends](https://newruntime.com/topics/trends.md): Explore the trends topic hub. - collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis. - collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph. - whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)