⏱️ Reading time: 8 min
Moonshot AI activated Kimi K3.1 Pro today on Kimi.ai, Kimi Work, Kimi Code, and its API, though the model’s open weights won’t arrive until November 4. In the meantime, it already competes closely with Claude Opus 5.5 in agentic coding: 65.7% versus 66.4% on Terminal-Bench 4.0.
📑 En este artículo
The release also adds K3.1 Pro Highspeed, a variant that’s 2 to 3 times faster, designed for real-time interactions. It’s the first release in the new K3.1 series and Moonshot AI’s biggest leap since Kimi K3.
TL;DR
- Moonshot AI launched Kimi K3.1 Pro today, October 23, 2026, with a 1 million token context window.
- The model scores 65.7% on Terminal-Bench 4.0, above GPT-6 Astra (58.2%) and close to Claude Opus 5.5 (66.4%).
- The full open weights will be released before November 4, 2026.
- The K3.1 Pro Highspeed variant offers 2 to 3 times faster response speed.
- In finance, the model improves 11.6 points over Kimi K3 on the Financial Research benchmark.
What Is Kimi K3.1 Pro
Kimi K3.1 Pro is the flagship open-source model Moonshot AI added to its new K3.1 series, with released weights and a 1 million token context window. It’s a post-trained version of Kimi K3 using large-scale reinforcement learning, optimized for coding, 3D modeling, and financial analysis.
Moonshot AI describes it as its most capable open model in the 3-trillion-parameter class to date. It’s available today across all four products in the Kimi family: the general chat assistant on Kimi.ai, the Kimi Work workspace, the Kimi Code programming assistant, and the developer API.
Unlike a from-scratch release, K3.1 Pro starts from the same base weights as Kimi K3 and adds a much larger layer of reinforcement learning, applied across hundreds of different domains at once.
What Happened
The announcement came with concrete performance figures. On Terminal-Bench 4.0, which evaluates code agents on terminal tasks with limited resources, Kimi K3.1 scores 65.7%. That puts it above GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%).
On the internal Kimi Code Bench 3.0 Long-horizon benchmark, which measures multi-step execution and tool use on engineering tasks, the model reaches 61.2%, up from 55.95% for Kimi K3. The nearly 5.3-point improvement comes from the same expanded reinforcement learning that defines the entire K3.1 series.
K3.1 Pro Highspeed runs 2 to 3 times faster than the base model, designed for chats and agents that need immediate responses. Moonshot AI hasn’t yet published separate benchmarks for that version, so its actual quality compared to the standard model remains independently unconfirmed.
Context and History
Moonshot AI is the Chinese lab behind the Kimi family, which went from being a single chat assistant to an ecosystem with separate versions for work, coding, and agents. That split arrived with Kimi K3, divided between Kimi Work, Kimi Code, and Kimi Claw.
K3.1 Pro is the direct successor to that model: it takes the same base weights and adds a much larger round of reinforcement learning across hundreds of domains at once. Moonshot AI positions it as its most direct answer yet to Anthropic’s and OpenAI’s closed models on code agent tasks.
Technical Details
The performance leap doesn’t come from a new architecture, but from how Moonshot AI organizes reinforcement learning. The company redesigned, layer by layer, the RL framework, long-context inference, training efficiency, sandbox infrastructure, and rollout planning, along with a shared platform that reviews the quality of each training environment.
That redesign allows each additional unit of RL compute to produce a stable, predictable gain in each domain, instead of improvements that stall in some areas while others regress. The following diagram summarizes that flow:
flowchart TD
A["RL Framework"] --> B["Long-context Inference"]
B --> C["Training Efficiency"]
C --> D["Sandbox Infrastructure"]
D --> E["Rollout Planning"]
E --> F["Environment Quality Check"]
F --> G["Stable Gain per RL Compute"]The most concrete case Moonshot AI documented is a human walking dynamics simulator, built on human-atlas, an open anatomical model with 296 bone components and 402 muscle components. Across 167 missions and 420 independent reviews, the model analyzed experimental results, diagnosed errors, and refined the gait controller until the simulation went from collapsing within seconds to walking for 11.2 seconds with alternating strides.
When adjusting the reward function stopped improving arm movement, the system didn’t keep pushing with more tweaks. It ran open-loop tests on the simulator’s physics, found and fixed 36 errors in the shoulder muscle lever-arm calculation, and only then retrained the controller.
In 3D modeling, K3.1 Pro uses spatial reasoning to infer shape, proportions, and relationships between parts from a sketch, a reference image, or a text description. On BenchCAD, which measures 3D reconstruction from multi-view renders through to generated CAD code, it scores 92.7% with tools, close to Claude Opus 5.5’s 96.2% and GPT-6 Astra’s 95.9%. It can write and run scripts for Blender and CAD software without relying on MCP connectors, useful for parametric modeling of mechanical parts or configurable furniture.
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 (code agents) | 65.7% | 66.4% | 58.2% |
| BenchCAD (3D modeling with tools) | 92.7% | 96.2% | 95.9% |
In finance, the model reaches 64.4 points on OfficeQA Pro v2, the best reported result for that benchmark, which requires analyzing real tax data with cross-document verification. It also improves 7.2 points over Kimi K3 on Financial Modeling, which evaluates formula links and financial logic in real spreadsheets, and 11.6 points on Financial Research, which measures whether the model produces auditable research under evidence and cutoff-date constraints. On Business Operation, a simulated long-term business management environment, the model performs about 4.2 times better than Kimi K3.
Impact and Analysis
The most relevant result isn’t that K3.1 Pro wins on any single benchmark, but that an open model lands within less than a point of a reference closed model on real agentic tasks. That weakens the argument that only closed labs can sustain this level of performance.
📌 Note: the open weights allow self-hosting K3.1 Pro, but a model in the 3-trillion-parameter class requires clusters with hundreds of gigabytes of GPU memory: it’s not something that runs on a laptop or a single server.
For LATAM teams, the 1 million token window matters as much as the score. It allows loading entire codebases, long financial documents, or extensive conversation histories without fragmenting the prompt. But real access, until the weights arrive on November 4, goes through the API or Kimi’s products, not through self-hosted infrastructure.
The competition is also moving fast: Claude Opus 5.5 remains slightly ahead on Terminal-Bench 4.0 and BenchCAD, and GPT-6 Astra trails behind on both. The difference between the three today is measured in percentage points, not full model generations.
What’s Next
The full weights for Kimi K3.1 arrive on November 4, 2026, which will let the community run independent evaluations and confirm whether Moonshot AI’s numbers hold up outside its own internal benchmarks. Until then, any comparison depends on the data the company itself published.
It also remains to be seen whether K3.1 Pro Highspeed maintains the base model’s quality at 2 to 3 times the speed, something Moonshot AI hasn’t yet backed up with its own benchmarks. And whether 3D Make, already live on kimi.ai, translates into real adoption beyond the examples the company itself showed.
Try it yourself: go to kimi.ai today and compare a long coding task against whatever model you’re currently using, before the open weights arrive on November 4.
Frequently Asked Questions
What is Kimi K3.1 Pro and who develops it?
It’s the flagship open-source model in Moonshot AI’s new K3.1 series, with a 1 million token context window and post-trained on Kimi K3 using large-scale reinforcement learning.
When will Kimi K3.1’s open weights be released?
Moonshot AI confirmed the model’s full weights will be available before November 4, 2026; today it’s only available through its products and API.
What sets K3.1 Pro Highspeed apart from the standard model?
It runs 2 to 3 times faster, designed for real-time interactions, though Moonshot AI hasn’t published its own benchmarks for that variant.
How does Moonshot AI’s model compare to Claude Opus 5.5?
It lands just behind on Terminal-Bench 4.0 (65.7% versus 66.4%) and BenchCAD (92.7% versus 96.2%), a difference of less than a point in the first case.
Can Kimi K3.1 be used for 3D modeling?
Yes: it scores 92.7% with tools on BenchCAD and can write scripts for Blender and CAD software without MCP connectors, as well as generate parametric models.
References
- Kimi (official blog): original Kimi K3.1 Pro announcement with benchmarks and availability.
- Moonshot AI’s GitHub: official repositories for the Kimi model family.
- Hugging Face: Moonshot AI: official organization where the open weights of the Kimi models are published.
📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Featured image: Foto de Numan Ali en Unsplash
Did it work for you? Got a different error? Say so below: questions get answered and help the next reader.
Leave a comment
0 Comments