On October 23, 2026, Moonshot AI released Kimi K3.1 Pro, the first model in their new K3.1 series and their most capable 3T-class open-weight model to date. This isn't just another incremental update—K3.1 Pro represents a substantial leap in what open-weight models can achieve, narrowing the gap between open and proprietary frontier models from 14 points (K3's 44 to Claude Opus 5.5's 58) to just 3 points (K3.1 Pro's 55 to 58) on the Artificial Analysis Intelligence Index.
The numbers tell a compelling story: K3.1 Pro scores 65.7% on Terminal-Bench 4.0 in Moonshot's evaluation (though independent testing by Artificial Analysis measured 67%, potentially the highest score of any model on this benchmark), surpassing GPT-6 Astra (58.2%) and narrowly trailing Claude Opus 5.5 (66.4%). It ranks 4th overall on the Intelligence Index with a score of 55, ahead of GPT-6 Astra (53), Gemini 4 Argon (53), and every other publicly available model except three Claude variants. On GDPval-AA, which measures performance on real-world work tasks, K3.1 Pro achieves 1,846 Elo—3rd place globally, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo).
What makes K3.1 Pro particularly significant is its open-weight release strategy. While proprietary models like Claude and GPT remain black boxes, Moonshot will release the full model weights by November 4, 2026. This makes K3.1 Pro the highest-scoring open-weight model currently available, giving researchers, enterprises, and developers unprecedented access to frontier-level AI capabilities.
This deep dive examines the technical architecture, training methodology, benchmark performance, cybersecurity controls, and competitive positioning of Kimi K3.1 Pro. We'll explore how Moonshot achieved these results through massive-scale reinforcement learning, what the Stable LatentMoE framework actually does, where K3.1 Pro excels and where it falls short, and what this release means for the future of open AI development.
Kimi K3.1 Pro is explicitly described as a post-trained version of Kimi K3 with substantially larger post-training compute. The base architecture remains identical—same 2.8 trillion total parameters, same 1-million-token context window, same Kimi Delta Attention and Attention Residuals, same Stable LatentMoE framework. What changed is the reinforcement learning phase.
According to Moonshot's technical blog, K3.1 Pro benefited from optimizations across every layer of the RL training stack:
The result: stable, predictable returns from each increment of RL compute. Moonshot reports that K3.1 Pro achieves consistent gains in every domain tested. On their internal Kimi Code Bench 3.0 Long-horizon, which tests coding agents' multi-step execution and tool use, K3.1 Pro scores 61.2% compared to K3's 55.95%—a 5.25 percentage point improvement.
Financial analysis and investment research demand high factual accuracy. Moonshot made targeted improvements to reduce hallucinations. On Kimi Hallu Bench Hard, an internal stress-test benchmark designed to probe situations where models are prone to errors:
This is a substantial reduction. Moonshot emphasizes that these are stress-test scenarios, not representative of overall error rates in daily use, but the improvement demonstrates effective post-training control over factual grounding.
As K3.1 Pro became more capable, Moonshot observed that cyber-related content was embedded within the general mid-training corpus, and cyber-related tasks were present in the RL environment mix. Without intervention, downstream cyber capability would increase substantially. For the open-weight release, Moonshot:
The checkpoint without these controls solved 90.5% of xdev-bench-hard problems at pass@3. The final K3.1 Pro reaches 81.0%. This 9.5 percentage point reduction demonstrates that targeted data removal meaningfully reduces offensive cyber capability while preserving general coding strength.
Kimi K3.1 Pro's architecture is inherited directly from K3. Here's what we know:
The config file shows hidden_size: 7168 and moe_intermediate_size: 3072. A naive calculation (92 layers × 896 experts × 3 × 7168 × 3072) yields approximately 5.45T parameters, but the actual download size is 1.42 TB, which at 4-bit quantization corresponds to roughly 2.7-2.8T parameters, consistent with Moonshot's claim.
Kimi Delta Attention is a modification of standard attention mechanisms designed to improve long-context handling. While Moonshot has not released full architectural details, KDA is described as enabling:
The exact mechanism remains proprietary, but the benchmark results on long-horizon tasks suggest KDA successfully addresses the quadratic complexity bottleneck of standard attention.
Attention Residuals add a skip connection around the attention mechanism itself, allowing gradients to flow more directly during backpropagation. This architectural choice likely contributes to training stability and enables deeper networks without vanishing gradients.
Unlike most transformers, which use RoPE (Rotary Position Embeddings) or absolute positional encodings, K3.1 Pro uses NoPE — No Positional Embeddings throughout the entire network. This design choice:
NoPE is an unconventional choice that reflects Moonshot's focus on extreme long-context capability. The tradeoff is that the model must learn positional patterns implicitly, which may require more training compute but offers greater flexibility at inference time.
Stable LatentMoE is Moonshot's approach to scaling MoE sparsity while maintaining training stability. Traditional MoE models suffer from:
Stable LatentMoE addresses these issues by:
The result is a model that activates only 16 out of 896 experts per token—an extraordinarily sparse activation rate of just 1.79%—while achieving frontier-level performance. This sparsity is what makes K3.1 Pro economically viable: 104B active parameters deliver compute comparable to models like GPT-4, but inference cost scales with active parameters, not total parameters.
K3.1 Pro accepts text, images, and video through the API. The model has native vision capabilities, meaning visual inputs are processed directly by the same transformer architecture rather than requiring a separate vision encoder. This unified approach enables:
Moonshot published extensive benchmark results across coding, agentic, reasoning, knowledge, multimodal, and cybersecurity domains. Here's where K3.1 Pro demonstrates frontier-level performance.
| Benchmark | K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra | DeepSeek V4.1 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% | 58.2% | — |
| Terminal-Bench Science 0.1 | 12.4% | — | — | 15.7% |
| DeepSWE v1.1 | 55.8% | 58.1% | 52.3% | 52.1% |
| SWE-Marathon v1.1 | 27.3% | 29.8% | 24.1% | — |
| PostTrainBench | 42.6% | — | — | 39.2% |
| ProgramBench (binary) | 81.3% | — | — | 78.9% |
| BenchCAD (with tools) | 92.7% | 96.2% | 95.9% | — |
Analysis: K3.1 Pro is competitive across the board, with particularly strong performance on Terminal-Bench 4.0. Moonshot reports 65.7%, which trails Claude Opus 5.5 (66.4%) by 0.7 percentage points. However, Artificial Analysis's independent testing measured 67% for K3.1 Pro—potentially the highest score of any model on this benchmark. On DeepSWE v1.1 and SWE-Marathon—benchmarks that test software engineering agents on real-world GitHub issues—K3.1 Pro trails Opus 5.5 but beats GPT-6 Astra and DeepSeek V4.1.
| Benchmark | K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| GDPval-AA v2.1 | 1,846 Elo (3rd) | 1,866 Elo (1st) | 1,803 Elo |
| BioMysteryBench | 68.2% | 71.4% | — |
| Our Agents' Last Exam | 53.3% | — | 58.2% |
| OfficeQA Pro v2 | 64.4 | — | — |
| HANDBOOK.md | 67.8% | — | — |
| Harvey Legal Agent | 72.1% | — | — |
| OSWorld V2.1 | 48.3% | — | 52.7% |
| MCP Atlas (500-task subset) | 81.2% | — | — |
| BrowseComp | 58.4% | — | — |
Analysis: The standout result is GDPval-AA, where K3.1 Pro ranks 3rd globally with 1,846 Elo—a massive 312 Elo gain over K3 (1,534 Elo). This benchmark measures performance on real-world work tasks, and K3.1 Pro trails only Claude Opus 5.5 and Muse Wonder Full.
On OfficeQA Pro v2, which tests complex fiscal data analysis over real financial documents through retrieval, cross-document verification, and multi-step numerical reasoning, K3.1 Pro achieves state-of-the-art results with a score of 64.4. This aligns with Moonshot's emphasis on financial research and modeling as a core use case.
| Benchmark | K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| HLE (2,500 questions) | 78.3% | — | — |
| HLE-Diamond (1,000 questions) | 72.1% | — | — |
| GDP.pdf (strict pass rate) | 61.8% | — | — |
| MMMU-Pro | 54.7% | — | — |
Analysis: K3.1 Pro demonstrates strong general knowledge and reasoning capabilities. HLE and HLE-Diamond test factual knowledge and multi-hop reasoning across diverse domains. GDP.pdf evaluates the model's ability to process complex PDF documents and answer rubric-based questions accurately.
| Benchmark | K3.1 Pro | Notes |
|---|---|---|
| GDP.pdf w/ Python | 68.4% | Task-macro mean rubric pass rate (5 runs) |
| ZeroBench w/ Python | 72.9% | Averaged over 5 runs |
| ChartoGraphy w/ Python | 81.3% | Averaged over 5 runs |
| BenchCAD (without tools) | 87.6% | 3D reconstruction from multi-view renderings |
Analysis: K3.1 Pro excels at multimodal tasks that combine vision and code execution. On BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, K3.1 Pro scores 92.7% with tools and 87.6% without tools, approaching the performance of Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).
| Benchmark | K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| ExploitBench (pass@5 max) | 12.8/16 | — | 14.1/16 |
| ExploitGym (6-hour budget) | 73.2% | 78.4% | 76.9% |
| xdev-bench-hard (pass@3) | 81.0% | — | — |
Analysis: K3.1 Pro's offensive cyber capability is strong but deliberately constrained. On xdev-bench-hard, the model reaches 81.0% after cyber-related data removal, compared to 90.5% for an uncontrolled checkpoint. Moonshot emphasizes that K3.1 Pro's cyber capability remains behind frontier proprietary models, aligning with their open-weight release principles.
Beyond benchmarks, Moonshot showcased several real-world applications where K3.1 Pro demonstrates frontier-level performance.
In one internal test, K3.1 Pro built a human walking dynamics simulator with integrated training and evaluation pipelines on top of human-atlas, an open-source anatomical model comprising 296 skeletal components and 402 muscle components. Across 167 missions and 420 independent reviews, the model:
When reward shaping failed to improve arm motion, K3.1 Pro didn't simply keep adjusting the reward function. Instead, it:
This demonstrates sophisticated debugging and hypothesis testing—hallmarks of expert-level engineering work.
K3.1 Pro excels at 3D creation and spatial reasoning. From a text brief, sketch, or reference image, it can infer shape, proportions, and spatial relationships to create usable 3D models. Moonshot highlights parametric modeling as a particular strength:
Without MCP connectors, K3.1 Pro can write and run scripts for tools like Blender and CAD software to build and refine assets, edit components, and adjust geometry, materials, and scene layout. Moonshot launched 3D Make on kimi.ai and the Kimi app, allowing users to preview and edit 3D models online, download them ready for 3D printing, or export them as videos or Live Photos.
K3.1 Pro shows significant improvements across Moonshot's internal finance benchmarks:
On OfficeQA Pro v2, K3.1 Pro achieves state-of-the-art results with a score of 64.4, demonstrating its ability to handle complex fiscal data analysis over real financial documents through retrieval, cross-document verification, and multi-step numerical reasoning.
Kimi K3.1 Pro's API pricing is half the cost of Claude Opus 5.5:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Kimi K3.1 Pro | $2 | $10 |
| Kimi K3.1 Pro Highspeed | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
| GPT-6.1 Sol | $2 | $10 |
| GPT-6 Astra | $2.50 | $10 |
| Kimi K3 | $3 | $15 |
Kimi K3.1 Pro Highspeed delivers 2-3× the speed of K3.1 Pro for more responsive interactions, at the same price. This variant likely uses more aggressive KV cache optimization or shorter maximum thinking time, trading some quality for latency.
Kimi K3.1 Pro is available today through multiple channels:
/model commandOpen-Weight Release: Full model weights will be released by November 4, 2026. This makes K3.1 Pro the highest-scoring open-weight model on the Artificial Analysis Intelligence Index.
Terminal-Bench 4.0 evaluates coding agents on realistic, resource-constrained terminal tasks using strict task-specific tests. This is one of the most demanding benchmarks for coding agents:
Using Moonshot's reported figure, K3.1 Pro trails Opus 5.5 by just 0.7 percentage points while beating GPT-6 Astra by 7.5 percentage points. Artificial Analysis's independent measurement of 67% would place K3.1 Pro at the top of this benchmark, ahead of all other models tested. This positions K3.1 Pro as a near-frontier (or frontier-leading) coding agent at half the price of Opus 5.5.
GDPval-AA measures performance on real-world work tasks using an Elo rating system:
K3.1 Pro trails the leader by just 20 Elo and beats GPT-6 Astra by 43 Elo. The gap between K3.1 Pro and the frontier is narrower here than on any other agentic benchmark.
BenchCAD tests 3D reconstruction through CAD code generation:
K3.1 Pro trails by 3.5 percentage points, demonstrating strong spatial reasoning and code generation capabilities, though not quite matching the frontier proprietary models.
At $2 input / $10 output per million tokens, K3.1 Pro costs exactly half as much as Claude Opus 5.5 ($4/$20) and matches GPT-6.1 Sol pricing. For enterprises running high-throughput agentic workflows, this price difference is significant. A workload generating 100 million output tokens per month would cost:
If K3.1 Pro delivers 95% of Opus 5.5's performance at 50% of the cost, the cost-effectiveness proposition is compelling.
Despite its impressive performance, K3.1 Pro has notable gaps compared to frontier proprietary models:
On software engineering benchmarks that test agents on real-world GitHub issues:
While K3.1 Pro beats GPT-6 Astra and DeepSeek V4.1 on these benchmarks, it still trails Claude Opus 5.5. The gap suggests that Anthropic's RL training includes more diverse software engineering environments or more effective exploration strategies.
On the scientific subset of Terminal-Bench:
K3.1 Pro trails DeepSeek V4.1 by 3.3 percentage points on scientific computing tasks. This may reflect differences in mid-training data or RL environment coverage for scientific domains.
OSWorld evaluates desktop environment automation:
K3.1 Pro trails GPT-6 Astra by 4.4 percentage points on GUI automation tasks. This suggests OpenAI's RL training may include more extensive GUI interaction environments.
While Moonshot will release the full model weights, they have not published:
This lack of transparency contrasts with truly open models like Bonsai 2 27B, where full architectural and training details are available.
Moonshot describes K3.1 Pro as an "open-weight model" and will release the full weights by November 4, 2026. But what does "open-weight" actually mean, and how does it differ from "open-source"?
K3.1 Pro falls into the open-weight category. Moonshot will release the weights, but has not committed to releasing:
Even without full reproducibility, open-weight releases provide significant value:
For enterprises concerned about data privacy or regulatory compliance, the ability to run K3.1 Pro on-premises is a decisive advantage over proprietary models.
Moonshot has not yet disclosed the license terms for K3.1 Pro's weights. The K3 release used a custom license that allowed:
We expect K3.1 Pro to use a similar license, but confirmation will come with the November 4 weight release.
Moonshot's approach to cybersecurity capability in K3.1 Pro reflects a nuanced philosophy about open-weight releases and dual-use capabilities.
As models become more capable at coding, they naturally develop offensive cybersecurity capabilities—finding vulnerabilities, writing exploits, and understanding attack vectors. This creates a tension:
Proprietary models like GPT-6 and Claude can gate access through usage policies, content filters, and account controls. Open-weight models cannot—once the weights are released, anyone can use them for any purpose.
Moonshot observed that cyber capability was increasing as a natural byproduct of training on broad coding data. They responded by:
The result: K3.1 Pro's offensive cyber capability is meaningfully reduced (81.0% vs. 90.5% for an uncontrolled checkpoint) but still substantially above its predecessor K3 due to improved general coding and debugging capabilities.
Moonshot articulates two guiding principles:
This approach acknowledges that offensive cybersecurity capability is a form of dual-use technology—valuable for defense but risky if concentrated. By making defensive capabilities available to more teams, especially those maintaining legacy systems, Moonshot believes the net effect is positive for software security.
| Model | xdev-bench-hard (pass@3) |
|---|---|
| K3.1 Pro (uncontrolled checkpoint) | 90.5% |
| K3.1 Pro (final release) | 81.0% |
| GLM-5.3 | — |
| DeepSeek V4.1 | — |
The 9.5 percentage point reduction demonstrates that targeted data removal works. The model retains strong general coding ability (evidenced by its Terminal-Bench performance) but is less effective at the specific task of vulnerability exploitation.
K3.1 Pro supports a 1-million-token context window, but Moonshot has not disclosed:
Users should test K3.1 Pro's behavior at various context lengths before relying on the full 1M-token window in production.
Kimi K3.1 Pro Highspeed delivers 2-3× the speed at the same price. Moonshot has not disclosed the technical mechanism, but possibilities include:
The quality-latency tradeoff likely depends on the task. For simple queries, Highspeed may be indistinguishable from the base model. For complex reasoning tasks, quality may suffer.
K3.1 Pro accepts text, images, and video, but architectural details are sparse:
These details matter for users planning multimodal workflows.
With 2.8T total parameters and 104B active parameters, K3.1 Pro requires substantial hardware:
Moonshot has not yet released official deployment guides or hardware recommendations. Users should expect significant infrastructure requirements for self-hosting.
K3.1 Pro demonstrates that open-weight models can now compete directly with the best proprietary systems. The gap on the Artificial Analysis Intelligence Index has narrowed from 14 points (K3's 44 vs. Claude Opus 5.5's 58) to just 3 points (K3.1 Pro's 55 vs. 58). On specific benchmarks like Terminal-Bench 4.0 and GDPval-AA, the gap is even smaller.
This has profound implications:
K3.1 Pro is the strongest model out of China to date, ahead of GLM-5.3 and DeepSeek V4.1 on most benchmarks. This reflects the maturation of China's AI ecosystem:
Moonshot, DeepSeek, Zhipu AI (GLM), and others are no longer playing catch-up—they're competing at the frontier.
At $2/$10 per million tokens, K3.1 Pro is half the price of Claude Opus 5.5 with comparable performance. This puts pricing pressure on proprietary model providers:
Moonshot's deliberate reduction of offensive cyber capability in K3.1 Pro sets a precedent for responsible open-weight releases of dual-use capabilities. Other labs will face similar decisions as models become more capable. Key questions:
Moonshot's approach—don't extend the absolute frontier, but progressively raise the floor—offers one answer, but the debate will continue.
K3.1 Pro is a strong choice for:
Consider Claude Opus 5.5 if:
Consider GPT-6 Astra if:
Consider smaller models (Claude Sonnet 5.5, GPT-6.1 Sol) if:
Before committing to K3.1 Pro for production use:
Kimi K3.1 Pro marks an inflection point in AI development. For the first time, an open-weight model delivers frontier-level performance across coding, agentic, and knowledge tasks—at half the price of the best proprietary alternatives.
The technical achievements are substantial:
But K3.1 Pro's significance extends beyond raw performance. By releasing frontier capabilities as open weights, Moonshot is accelerating a shift toward a more open, accessible, and distributed AI ecosystem. Enterprises gain more control, researchers gain more transparency, and developers gain more freedom to innovate.
The gap between open and proprietary models has narrowed dramatically—from 12 points to 3 on the Intelligence Index. If this trend continues, the distinction between "open" and "proprietary" may soon lose its meaning. The frontier will simply be the frontier, regardless of whether the weights are released.
For practitioners, the message is clear: open-weight models are now viable alternatives to proprietary systems for production workloads. K3.1 Pro won't be the right choice for every use case, but it's competitive enough that you should evaluate it seriously.
The AI landscape has entered a new equilibrium. The best models are still proprietary, but the gap is closing fast—and the open-weight frontier is real.