Kimi K3.1 Pro: 2.8-trillion-parameter model goes open weights on November 4

What this is about

Chinese AI company Moonshot has unveiled Kimi K3.1 Pro. It's a post-trained version of Kimi K3 that used far more compute for reinforcement learning (learning through reward) during post-training. The model is multimodal and built as a mixture-of-experts (MoE): of its 2.8 trillion parameters, roughly the dials tuned during training, only 104 billion are active per token. The context window is 1 million tokens. At launch, the model is available through the Kimi website and app, Kimi Work, Kimi Code and the API. The weights aren't available yet, but Moonshot has announced them for November 4, 2026.

The AINews newsletter (Latent Space) frames this as narrowing the gap between open and closed models. According to Artificial Analysis (AA), once the weights are released, the Intelligence Index gap between the best open and the best proprietary model would shrink from 12 points to 3.

What Moonshot promises

Moonshot touts complete projects in Unreal, Unity and Godot, parametric modeling in Blender and CAD, financial modeling and long-running agents. There are also "Kimi Hosted Agents," a managed environment that runs around the clock and lets multiple agents work together. Moonshot cites 65% fewer hallucinations on finance tasks. The source doesn't give the baseline, denominator or protocol behind that number, so treat it as a vendor claim.

Little is known about the architecture: layout, expert routing, attention design and training budget haven't been published. The newsletter calculates, for illustration, that 2.8 trillion parameters alone come to about 5.6 TB at 16-bit or 1.4 TB at 4-bit, before overhead. To run the model, you need to keep all the experts available, not just the 104 billion active ones.

Independent results

• Artificial Analysis: 4th place with 55 points. The gains are mainly in agentic tasks. On AA-Omniscience, accuracy rose from 48% to 51%, and the hallucination rate fell from 53% to 44%.
• Arena (Code Arena WebDev): 8th place with 1,692 points, nine places and 37 points above K3. After release, Arena says it would be the top-ranked open model, overtaking DeepSeek V4.1 Flash. It sits roughly level with Opus 5 Max, not the newer Opus 5.5.
• Aggregate ECI: a preliminary 164.7 with a wide interval (152.5–169.0). That's an estimate from published benchmarks, not a measurement of its own.

The picture is more mixed for 3D and interactive tasks: early AA 3D votes point to a lower rank than in coding. On the RuneScape bench, one user says it landed roughly on par with DeepSeek V4.1 Flash and Opus 5.0. Another user reports that on a compiler project, neither K3.1 nor its alternatives matched Fable 5.1 or Opus 5.5. These are individual cases, not overall results.

Pricing and actual cost

High-speed costs double and is said to be 2–3× faster. But cheaper tokens don't automatically mean cheaper tasks. In the AA index, a task costs $1.98 (roughly unchanged) because K3.1 Pro produces 55,000 output tokens instead of 48,000. On the AA Cyber Index, it's $2.81 per task (15% more). In the Arena WebDev measurement, the median is $1.23: 29% below K3, 74% below Opus 5 Max and 24% below GPT-6 Sol Max.

Cybersecurity

Moonshot removed cyber-related data from mid-training and RL, rather than relying only on refusals at runtime. An internal metric reportedly fell from 90.5% to 81%. Even so, the model scores 64% on the AA Cyber Index (K3: 41%, mainly CyberGym, up from 58% to 95%). That's the highest publicly available result. Only GPT-6 Sol Daybreak Blue, which has special access, does better at 69%. AA observed no refusals. Whether that reflects a safety measure or simply reduced capability remains open. Remarks like "annoyed" and comparisons with other models are commentary, not findings.

Criticism and controversies

• Benchmarks: Moonshot's table isn't an independent replication. Critics point to the use of the Claude Code harness and inconsistent sources for Terminal-Bench comparisons.
• Running costs: One commenter estimates the model is 30× or more expensive to run than DeepSeek Flash and called for a Flash model. These are unsupported estimates. No Flash model has been announced.
• "Local": A trillion-parameter model can be self-hosted, but it won't run on consumer devices.
• Provider differences: For K3 (not K3.1), Vals found 60.6% at Baseten versus 75.8% at Moonshot on legal tasks.
• Paid promotion: Several users claim influencers were allegedly barred from adding a paid-partnership label. The briefing isn't authenticated, and there's no response from Moonshot.

Takeaway

If the weights arrive as announced, this would be an open model close to the frontier, which observers had only expected by year's end. The license, checkpoint precision and hardware requirements are still open, so "open weights soon" is more accurate than "open source." For most readers, the API matters more than self-hosting. The independent tests from AA and Arena are more reliable than vendor tables and influencer opinions, but they also show some weaknesses in 3D and in cost per task.