Quick Answer: Kimi K3.1 Pro is Kimi's post-trained successor to K3, with a 1-million-token context window and a faster Highspeed variant that Kimi says runs 2–3× quicker. Kimi's own numbers put it near the top of coding, 3D, and finance benchmarks, but those results are vendor-reported. The full open weights are due by November 4, 2026.
Most model launches lead with one leaderboard number. Kimi's K3.1 Pro announcement does something rarer: it reports a benchmark where it loses, a hallucination rate it cut by more than half, and a cyber capability it deliberately made weaker. That mix is worth reading closely. Below, I'll walk through what Kimi actually shipped, what the numbers mean, and where you should be skeptical before you build on it.
What Kimi Actually Shipped
Kimi released two models. Kimi K3.1 Pro is the main model, described as the most capable 3T-class open-weight model Kimi has built so far. Kimi K3.1 Pro Highspeed targets responsiveness. Kimi says it delivers 2–3× the speed of the standard model, which matters more for interactive coding sessions than for long batch jobs.
Both models are live today on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The weights aren't out yet, though. Kimi says they'll arrive by November 4, 2026, which is about 11 days after this writing. So if you need to self-host, you'll have to wait.
Getting access depends on where you work:
- Kimi Code: Run Kimi Code and pick K3.1 Pro with the
/modelcommand. Kimi's model documentation lists the available options. - Kimi API: Use the model IDs
kimi-k3.1-proorkimi-k3.1-pro-highspeed. Zero Data Retention is available for API customers. - Kimi Work: Install version 3.1.0 or later of the desktop app. It's available for Windows and Apple silicon Macs.
- Enterprise: Administrators have to enable K3.1 Pro for their organization. It's off by default at launch.
Here's the catch for teams: the context window is large, but enterprise access is opt-in. If your security team reviews every model before rollout, plan for that approval step now.
Coding Gains, and Where They Stop
Kimi's headline coding claim is that K3.1 Pro "matches or approaches" the strongest proprietary models. On Terminal-Bench 4.0, which tests coding agents on resource-constrained terminal tasks, it scores 65.7%. That's ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. Kimi says it ran its model with the Claude Code harness and took the comparison scores from vendor blogs or the official leaderboard.
That's the number to remember: a 0.7-point gap. Most practitioners would call that a tie, and the honest reading is that K3.1 Pro is competitive with Opus 5.5 on this test, not better.
The more interesting result is Kimi's internal Kimi Code Bench 3.0 Long-horizon, which measures multi-step execution, tool use, and end-to-end problem solving. K3.1 Pro scores 61.2%, up from K3's 55.95%. Since Kimi wrote this benchmark, treat it as a signal about direction rather than a neutral ranking.
The case study Kimi highlights shows what long-horizon work looks like in practice. Kimi reports that the model built a human walking dynamics simulator on the open-source human-atlas anatomical model. Over 167 missions and 420 independent reviews, the simulated figure went from collapsing within seconds to walking with alternating strides for 11.2 seconds.
The detail that matters most: when reward shaping stopped improving arm motion, the model ran open-loop probes of the simulator's dynamics and found 36 errors in shoulder-muscle moment-arm calculations. Most fine-tuning workflows would have kept tweaking the reward function. Debugging the environment itself is the behavior you actually want from an agent, and it's the part most evaluations never measure.
3D Modeling and CAD Results
K3.1 Pro's 3D claims are the most specific in the announcement. On BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, it scores 92.7% with tools enabled. That's close to Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%, but it's still a gap of more than three points.
Kimi says the model is strongest at parametric work: mechanical parts, architectural models, configurable products, and custom furniture. Without MCP connectors, it can write and run scripts for Blender and CAD software to edit components and adjust geometry, materials, and layout. That's a useful distinction. A model that writes a Blender script you can read and rerun is much easier to audit than one that acts through a black-box plugin.
Kimi also launched 3D Make on kimi.ai and the Kimi app. You can preview and edit models in the browser, download them for 3D printing, or export videos and Live Photos.
Here's where most teams trip up: a high BenchCAD score doesn't tell you how the model handles your tolerances. Reconstructing a shape from renders is a different job from producing a part that fits a real assembly. Before you trust output for printing or manufacturing, check dimensions against your own specs.
Finance and the Hallucination Numbers
Finance is where Kimi makes its strongest claims. K3.1 Pro reports a state-of-the-art score of 64.4 on OfficeQA Pro v2, which tests retrieval, cross-document verification, and multi-step numerical reasoning over real financial documents. Kimi also reports gains over K3 on its internal finance benchmarks:
- Financial Modeling (formula linkage, financial logic, and workbook quality): up 7.2 points.
- Financial Research (auditable research under evidence and cutoff constraints): up 11.6 points.
- Business Operation (long-horizon decisions in a simulated business): about 4.2× K3's score.
The more useful figure for analysts is factual accuracy. On Kimi's Hallu Bench Hard, the claim-level hallucination rate dropped from 23.4% for K3 to 8.3%. The share of responses containing at least one factual error came in at 42.4% of K3's rate.
That last number is easy to misread, so the mechanics matter. Those rates come from a stress test that Kimi built with examples crafted to trigger errors. Kimi says outright that the figures don't represent error rates in everyday use. Your real error rate depends on your prompts, your documents, and whether the model has access to them.
This is where a counter-intuitive point comes in. Lower hallucination on a stress test can make the model look more trustworthy than it is for tasks it hasn't seen. In financial research, the failure mode that hurts is a confident, well-formatted wrong answer. Make the model cite sources for every number, and check those citations yourself.
The Cyber Trade-off Nobody Should Skip
This section is the one most coverage will skip, and it's the most unusual. Kimi reports that vulnerability research performance rose steadily as post-training scaled. The team traced the cause to cyber-related content in the general mid-training corpus and cyber tasks in the RL environment mix.
For the open-weight release, Kimi says it removed cyber-related data from both training stages while keeping general coding tasks. The effect is measurable. On xdev-bench-hard, an in-house benchmark of source-code vulnerability discovery and exploitation, the checkpoint without these measures solved 90.5% of problems at pass@3. The final K3.1 Pro solved 81.0%. Kimi says K3.1 Pro still improves substantially on its predecessor. Its offensive capability also remains behind frontier proprietary models that have been public for at least three months.
Kimi's argument has two parts. It wants to avoid extending the frontier of offensive cyber capability with open weights. It also wants to release open weights gradually, so that defenders can audit legacy code without paying for a dedicated security review.
Both sides have a real case. Teams that maintain old systems often can't afford specialist audits, so a capable local model is a practical defensive tool. On the other hand, a self-reported benchmark can't prove that the removed data was the only path to offensive skill. Treat the 9.5-point drop as Kimi's measurement, not an independent verdict.
Reading the Scorecard Before You Trust It
Here are the headline numbers side by side. Every row is reported by Kimi, and comparison scores come from the vendors' own posts or public leaderboards.
| Benchmark | K3.1 Pro | Comparison point | What it measures |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | Claude Opus 5.5: 66.4%; GPT-6 Astra: 58.2% | Coding agents on terminal tasks |
| BenchCAD (with tools) | 92.7% | Claude Opus 5.5: 96.2%; GPT-6 Astra: 95.9% | 3D reconstruction via CAD code |
| Kimi Code Bench 3.0 Long-horizon | 61.2% | K3: 55.95% | Multi-step agentic coding (Kimi internal) |
| xdev-bench-hard, pass@3 | 81.0% | Predecessor without cyber data removal: 90.5% | Vulnerability exploitation (Kimi internal) |
The pattern is clear. K3.1 Pro is strongest against its own predecessor and tends to trail the top proprietary models by small margins. That's a respectable place to be, especially for open weights. But a few cautions apply. Terminal-Bench and BenchCAD are public benchmarks, but the configurations and harnesses differ across vendors. The Kimi-internal benchmarks are designed by Kimi, which makes them useful for tracking progress and poor for ranking.
Run your own tests. Pick ten tasks you've already solved, rerun them on K3.1 Pro and your current model, and compare results. You'll learn more from that than from any table.
Frequently Asked Questions
What is Kimi K3.1 Pro?
Kimi K3.1 Pro is Kimi's post-trained successor to K3, released with a 1-million-token context window. Kimi positions it for coding, 3D creation, and financial research. A faster variant, Kimi K3.1 Pro Highspeed, claims 2–3× the speed. Both are available now on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
How do I use Kimi K3.1 Pro in Kimi Code?
Run Kimi Code and select the model with the /model command. If you're on an enterprise account, an administrator must first enable K3.1 Pro for your organization, because access is off by default. For direct API use, the model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed.
Is Kimi K3.1 Pro open source yet?
Not yet. Kimi says the full model weights will be released by November 4, 2026. Until then, you can only use it through Kimi's hosted products and API. Check the weights' license terms before commercial deployment, since "open-weight" doesn't always mean unrestricted use.
Why did Kimi remove cyber data from K3.1 Pro?
Kimi found that cyber-related content in its training data was a natural source of offensive capability. Removing it cut the pass@3 score on xdev-bench-hard from 90.5% to 81.0%. Kimi's stated goal is to limit how much offensive capability open weights can deliver, while still giving defenders useful tools.
The Bottom Line for Your Next Model Decision
K3.1 Pro is a serious open-weight option with strong vendor-reported gains in coding, 3D, and finance, and it's already usable through Kimi Code and the API. The open-weight release is still pending, and most of the numbers come from Kimi itself. So the sensible move is to test it on your own work before changing your stack. Open Kimi Code this week, switch to K3.1 Pro with /model, run one task that your current model keeps getting wrong, and write down the result. Then pass this along to someone who's tired of trusting launch-day benchmarks, because the real answer on Kimi K3.1 Pro will come from your own tasks.