Moonshot releases Kimi K3.1 Pro: 2.8 trillion parameters, weights due November 4

Moonshot has released Kimi K3.1 Pro, a multimodal MoE model with 2.8 trillion parameters, of which 104 billion are active, and a 1 million-token context window. The API and apps work from day one, and the company promises open weights on November 4, 2026. Artificial Analysis ranks the model fourth in its Intelligence Index with 55 points. If the scores hold up once the weights are published, the gap between the best open and best closed model on this index shrinks from 12 points to three.

K3.1 Pro is not a new architecture but K3 with a much larger compute budget for post-training. Moonshot rebuilt the whole stack: the RL framework, long-context inference, sandboxes, rollout scheduling, and a shared platform for verifying environment quality. That made it possible to run joint RL across hundreds of applied domains at once. The most visible gains are in agentic tasks; on knowledge and science tests they are more modest.

Where the model ranks

• Artificial Analysis Intelligence Index: Opus 5.5 — 58, Sonnet 5.5 and Muse Wonder Full — 56 each, Kimi K3.1 Pro — 55, fourth place.
• AA-Omniscience: accuracy rose from 48% to 51%, and the hallucination rate fell from 53% to 44%. That is nine percentage points; Moonshot's claimed "65% fewer hallucinations" refers to financial tasks and was measured differently.
• Code Arena WebDev: 8th place, 1 692 points. That is nine places and 37 points above K3. Brand & Marketing — 7th place, Simulations — 9th (previously 20th). Once the weights are published, it will be the highest-ranked open model on Arena, ahead of DeepSeek V4.1 Flash.
• Cyber Index: 64%, the best result among publicly available models and second overall after GPT-6 Sol Daybreak Blue (69%, trusted users only). The index rose 23 points from K3's 41%, and CyberGym jumped from 58% to 95%.

Token price and task price

Standard pricing: $2 per million input tokens, $10 for output tokens, and $0.20 for cached tokens. K3 cost $3/$15/$0.30, so this is a one-third discount. Opus 5.5 costs $4/$20, Astra — $10/$50. The Highspeed version costs twice the standard price and generates 2–3 times faster.

But you pay for completed work, not for tokens. On the Intelligence Index, K3.1 Pro produces 55 thousand output tokens per task versus 48 thousand for K3, 14% more, and that eats into part of the discount. The result is $1.98 per task, almost the same as its predecessor. On the cyber set, the model consumes roughly twice as many input tokens, so a task comes to $2.81, 15% more than K3 and noticeably more than DeepSeek 1017.

On Arena WebDev the picture is better: the median task costs $1.23, 29% less than K3. Comparison with other models:

• Opus 5 Max: 74% cheaper with a one-point preference advantage;
• GPT-6 Sol Max: 24% cheaper with an 11-point advantage;
• Qwen3.8-Max: 47% cheaper with a 20-point advantage.

Cybersecurity: the data was removed up front

Moonshot removed cyber data from mid-training and RL rather than relying only on refusals at inference time. According to the report, the internal score fell from 90.5% to 81%, while other capabilities kept improving. But the public API never refused a single defensive task in Artificial Analysis's tests. AA notes that closed models may have hidden capabilities above their public scores, as with GPT-6 Sol, whose trusted-access variant scored 32 points higher than the public one.

What to check before November 4

The main caveat concerns the word "local." 2.8 trillion parameters is roughly 5.6 TB at 16-bit or 1.4 TB at 4-bit, and that excludes quantization metadata, KV-cache, and tooling. The 104 billion active parameters speed up generation, but memory is needed for the entire pool of experts. Self-hosting requires an accelerator cluster, not a workstation.

In practice, this means the following. Teams doing agentic coding should run their own tasks and calculate the cost of a successful completion rather than rely on the price list. Those who care about latency can compare Highspeed at its doubled price. Anyone planning to self-host should wait for the license and commercial-use terms, which are not yet known. For a quick start, Cline already offers the model through ClinePass for $9.99 a month. Moonshot has also launched Kimi Hosted Agents, a managed environment with an always-on runtime for long-running tasks and multi-agent work.