Moonshot AI's Kimi K3.1 Pro Arrives at the Frontier — With Open Weights Promised by November 4

Moonshot AI's Kimi K3.1 Pro Arrives at the Frontier — With Open Weights Promised by November 4

Moonshot AI launched Kimi K3.1 Pro on October 23, 2026, delivering the most significant upgrade to its flagship model line since the original Kimi K3 debuted in July. The new model lands fourth on the Artificial Analysis Intelligence Index↗ with a score of 55, outperforming OpenAI's GPT-6 Astra (53) and Google's Gemini 4 Argon (53) — and trailing only the top three proprietary models from Anthropic. Alongside the launch, Moonshot cut API prices by 33% and committed to releasing the full open weights by November 4, 2026, a move that would make K3.1 Pro the highest-scoring open-weight model in existence.

The release arrives at a pivotal moment for Moonshot AI. The Beijing-based lab, founded in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin, has grown from a $2.5 billion startup to a company now valued at approximately $50 billion, with a Hong Kong IPO targeting Q1 2027 reportedly in preparation. Kimi K3.1 Pro is the technical statement that justifies that trajectory.

---

What Changed: Post-Training at Scale

Kimi K3.1 Pro is not a new pre-trained model. According to Moonshot's official blog post↗, it is a post-trained iteration of the original Kimi K3 architecture — the same 2.8 trillion total parameters and 104 billion active parameters per token via a Mixture-of-Experts (MoE) design. What changed is the volume and sophistication of post-training compute applied on top of that base.

The architecture itself is worth understanding. Kimi K3.1 Pro uses:

• 93 transformer layers with a hybrid attention design combining 69 Kimi Delta Attention (KDA) layers — a linear attention mechanism that reduces KV cache overhead over long contexts — and 24 Gated Multi-Head Latent Attention layers for global context snapshots
• 896 routed experts with 16 selected per token, plus two shared experts, enabling the model to specialize without activating its full parameter count
• NoPE (No Positional Embeddings) throughout the network, improving computational flexibility at long context lengths
• MXFP4 weights and MXFP8 activations via quantization-aware training, optimized for NVIDIA Blackwell and AMD MI400 accelerators
• A 1,048,576-token context window — one million tokens — maintained from the K3 baseline

The post-training investment is where the gains come from. Moonshot applied reinforcement learning across hundreds of domains, with particular emphasis on long-horizon coding, 3D creation workflows, and financial research. The result is a model that, on Artificial Analysis's independent evaluation↗, jumped from Kimi K3's Intelligence Index score of 44 to K3.1 Pro's 55 — an 11-point gain without touching the pre-training architecture.

"The gains are concentrated exactly where Moonshot aimed them: agentic coding, 3D scripting, and financial analysis. This is post-training as a precision instrument, not a blunt upgrade." — Artificial Analysis, October 23, 2026

---

Benchmark Breakdown: Where K3.1 Pro Leads and Where It Doesn't

The benchmark picture for Kimi K3.1 Pro is nuanced. Its strongest results are in agentic and coding tasks; its weakest are in closed-book scientific reasoning, where Anthropic and OpenAI's top models retain a clear edge.

Agentic and Coding Performance

• Terminal-Bench 4.0: K3.1 Pro scored 65.7%, placing it second among all models tested — trailing Claude Opus 5.5 (66.4%) but ahead of GPT-6 Astra (58.2%). Artificial Analysis's independent run put the figure slightly higher at 67%, the highest of any model in their evaluation suite.
• Kimi Code Bench 3.0 Long-horizon: Internal Moonshot evaluation showed 61.2%, up from 55.95% for Kimi K3 — a meaningful improvement in multi-file, environment-aware coding tasks.
• GDPval-AA (agentic benchmark): K3.1 Pro reached 1,846 Elo, a 312-point increase over Kimi K3, placing it third globally.
• AutomationBench-AA (SaaS workflow automation): Score rose from 58% to 73%, a 15-point jump that reflects the model's improved ability to navigate multi-step tool-use scenarios.

Specialized Domain Performance

• BenchCAD (3D modeling): K3.1 Pro achieved 92.7% with tools, approaching the performance of leading proprietary models on tasks involving Blender, CAD software, Unity, Unreal Engine, and Godot scripting.
• OfficeQA Pro v2 (financial research): Score of 64.4, with Moonshot reporting a 65% reduction in hallucinations during financial tasks compared to K3.
• Kimi Hallu Bench Hard (hallucination): Claim-level hallucination rate dropped from 23.4% (K3) to 8.3% (K3.1 Pro) — a dramatic improvement that matters enormously for enterprise deployments where factual accuracy is non-negotiable.

Where It Trails

Non-agentic reasoning benchmarks tell a more modest story. The AA-Omniscience Index rose from 20 to 30, and raw accuracy in closed-book scenarios sits at 51% — behind Claude Opus 5.5 and GPT-6 Astra in scientific reasoning tasks. K3.1 Pro is a specialist, not a generalist, and Moonshot's positioning reflects that honestly.

"K3.1 Pro is the strongest open-weight model for agentic coding tasks we have evaluated. The gap to Claude Opus 5.5 on Terminal-Bench is now under one percentage point." — Artificial Analysis Intelligence Index, October 23, 2026

---

Pricing: 33% Cut, Two Variants

Moonshot launched K3.1 Pro with a pricing structure that undercuts its predecessor while introducing a speed-optimized variant for latency-sensitive applications.

Standard K3.1 Pro

• Input: $2.00 per million tokens (down from $3.00 for K3)
• Output: $10.00 per million tokens (down from $15.00 for K3)
• Cache hit: $0.20 per million tokens

K3.1 Pro Highspeed

• Input: $4.00 per million tokens
• Output: $20.00 per million tokens
• Speed: 2–3× faster output generation, same underlying model

The Highspeed variant is priced at double the standard rate but targets real-time applications — interactive coding assistants, live financial analysis dashboards, and 3D modeling tools where latency directly affects user experience. Despite the higher per-token cost, Moonshot notes that total cost per Intelligence Index task remains approximately $1.98–$2.00 due to the model's improved efficiency per task.

Compared to Claude Opus 5.5 ($4.00 input / $20.00 output), the standard K3.1 Pro is half the price at comparable performance levels on agentic benchmarks. For developers running large-scale coding agents, that differential compounds quickly.

Access is available through the Kimi Open Platform API↗ using OpenAI-compatible endpoints with model identifiers `kimi-k3.1-pro` and `kimi-k3.1-pro-highspeed`. The model is also live on OpenRouter↗, Kimi.ai, Kimi Work (desktop version 3.1.0+), and Kimi Code. Enterprise accounts must manually enable the model through administrator settings.

---

The Open-Weight Commitment — and the Cybersecurity Tradeoff

The most consequential aspect of the K3.1 Pro launch may not be its benchmark scores but its open-weight commitment. Moonshot has promised to release the full model weights by November 4, 2026 — eleven days after launch. If delivered, K3.1 Pro would become the highest-scoring open-weight model on the Artificial Analysis Intelligence Index, surpassing the previous leader by a significant margin.

That commitment comes with a deliberate tradeoff. As The Next Web reported↗, Moonshot's researchers observed during development that scaling post-training compute naturally increased the model's vulnerability discovery and exploitation capabilities. Rather than release a model with frontier-level offensive cybersecurity potential, Moonshot made a specific intervention: removing cyber-related training data from both the mid-training corpus and the reinforcement learning environment.

The numbers tell the story clearly:

• Unredacted internal checkpoint: 90.5% pass@3 on xdev-bench-hard (a vulnerability research benchmark)
• Final K3.1 Pro (released): 81.0% pass@3 on the same benchmark

Moonshot frames the 9.5-point reduction as a deliberate safety margin — sufficient for defensive security auditing and general coding, but below the threshold that would meaningfully advance offensive capabilities. As WCCFTech noted↗, the final model still solves 81% of hard vulnerability research tasks, which is itself a formidable capability — the framing of "stripped" cybersecurity data somewhat obscures how capable the released model remains.

This approach mirrors the responsible disclosure framework that Moonshot applied to Kimi K3's open-weight release in July 2026, when the lab released weights one day ahead of schedule after completing safety evaluations. The pattern suggests Moonshot is developing a repeatable process for open-weight releases of frontier-class models — a capability that few labs globally have demonstrated.

---

Competitive Landscape: Where K3.1 Pro Fits

Against Chinese Peers

Within the Chinese AI ecosystem, K3.1 Pro immediately becomes the strongest publicly available model. DeepSeek's V4.1-Flash-1017 (released October 17) is a faster, cheaper model optimized for agentic coding at lower cost, but it does not compete at the same intelligence tier. Alibaba's Qwen 4 remains in training with no confirmed release date. Zhipu AI's GLM-5.3 (released August 14) is a strong cybersecurity-focused model but has not matched K3.1 Pro's agentic benchmark scores.

The competitive picture among Chinese labs as of October 23, 2026:

• Kimi K3.1 Pro (Moonshot AI): Intelligence Index 55, Terminal-Bench 65.7%, open weights promised Nov 4
• DeepSeek V4.1-Flash-1017: Optimized for cost-efficiency, MIT-licensed, strong coding but lower intelligence tier
• GLM-5.3 (Zhipu/Z.ai): 743B MoE, cybersecurity-specialized, staged open-weight rollout
• Qwen 4 (Alibaba): In training, no release date confirmed as of October 23

Against Western Frontier Models

The gap to the top proprietary models is narrowing but real. Claude Opus 5.5 leads the Artificial Analysis Intelligence Index at 58, with K3.1 Pro at 55 — a 3-point gap that translates to meaningful differences in scientific reasoning and closed-book accuracy. GPT-6 Astra sits at 53, meaning K3.1 Pro has already surpassed OpenAI's current frontier offering on the aggregate index.

The more significant comparison is on agentic tasks specifically. On Terminal-Bench 4.0, K3.1 Pro's 65.7% trails Claude Opus 5.5's 66.4% by less than one percentage point. For developers building coding agents, that gap is within noise — and K3.1 Pro costs half as much per token.

---

Moonshot AI's Trajectory: From Startup to $50 Billion

The K3.1 Pro launch is inseparable from Moonshot AI's broader corporate trajectory. The company's valuation has grown at a pace that few AI labs anywhere have matched:

• Early 2024: $2.5 billion valuation
• May 2026: $20 billion after a $2 billion raise
• July 2026: $35 billion after a $3.5 billion round
• October 2026: Approximately $50 billion following a final private round

Annual recurring revenue has reached approximately $1 billion, with projections toward $2 billion by December 2026. A Hong Kong IPO targeting Q1 2027 is reportedly in preparation, with Bloomberg reporting↗ that Bank of America, China International Capital, Deutsche Bank, and Goldman Sachs are engaged as sponsor banks. The company's backers include Alibaba Group, Tencent Holdings, 5Y Capital, IDG Capital, and the National Artificial Intelligence Industry Investment Fund.

K3.1 Pro is the technical proof point that Moonshot needs to sustain that valuation heading into a public offering. A model that places fourth globally on the Artificial Analysis Intelligence Index — ahead of GPT-6 Astra — while committing to open weights is a credible claim to frontier status.

---

How to Access Kimi K3.1 Pro

For developers ready to evaluate the model today:

• API access: Sign up at platform.kimi.ai↗ and use model ID `kimi-k3.1-pro` or `kimi-k3.1-pro-highspeed` via OpenAI-compatible endpoints
• Third-party routing: Available on OpenRouter↗ and AIML API for developers already using aggregator layers
• Consumer apps: Live on Kimi.ai, Kimi mobile apps, and Kimi Work desktop (version 3.1.0+)
• Kimi Code: Available for long-horizon coding workflows with video input support
• Open weights: Scheduled for public release by November 4, 2026 — enabling self-hosted deployment on multi-node GPU clusters

The reasoning effort setting (`low`, `high`, `max`) allows developers to trade cost against quality. At `low` effort, K3.1 Pro can complete coding agent tasks for roughly half the cost of Claude Opus 5.5, though with higher call volume and reasoning overhead. The `max` setting (default) delivers the benchmark scores cited above.

Kimi K3.1 Pro represents the clearest evidence yet that the gap between Chinese open-weight models and Western proprietary frontier systems has closed to within measurement error on the tasks that matter most to developers. Whether the open-weight release on November 4 delivers on that promise — and whether Moonshot's IPO trajectory holds — will define the next chapter for one of China's most closely watched AI labs.