Kimi K3.1 Pro ships: 2.8T-A104B, a 1M-token context and $2 per million input tokens
Moonshot AI released Kimi K3.1 Pro on Oct 23, the K3 series' first model — a post-trained K3 with 2.8T total / 104B active parameters, a 1M-token context and native multimodality. It scores 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2%. Pricing is $2.00 input / $10.00 output per million tokens, about 33% below the previous K3, plus a 2-3x faster Highspeed build at $4 / $20. Weights are promised by Nov 4, so the site logs it under the RESTRICTED convention with zero runs.
Source: araoai
Kimi K3.1 Pro ships: 2.8T-A104B, a 1M-token context and $2 per million input tokens
On Oct 23, Moonshot AI released Kimi K3.1 Pro, the first model in the K3 series and a post-trained version of K3: billed as the company's most capable 3T-class open-weight model to date, it keeps K3's 2.8T total / 104B active parameters, a 1,048,576-token (1M) context window and native multimodality (text/image/video). Community attention matched: the thread took 713 points and 453 comments on HN, the hottest LLM story of the window.
Benchmarks: 65.7% on Terminal-Bench 4.0
| Benchmark | K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | GPT-6 Astra 58.2% / Claude Opus 5.5 66.4% |
| Kimi Code Bench 3.0 Long-horizon | 61.2% | previous K3 55.95% |
| BenchCAD (with tools) | 92.7% | Claude Opus 5.5 96.2% / GPT-6 Astra 95.9% |
| OfficeQA Pro v2 | 64.4 | claimed state of the art |
Post-training is the point of this release: Moonshot describes it as large-scale joint RL across hundreds of downstream domains, which meant rewriting the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling and a shared environment-quality platform. 3D creation and finance improved in step — on BenchCAD it writes Blender and CAD scripts, and OfficeQA Pro v2 tests document-level fiscal reasoning.
Pricing: about 33% below the previous generation
Official API pricing per million tokens:
| Model | Input | Output | Cached input | Cache write 5m / 1h |
|---|---|---|---|---|
| kimi-k3.1-pro | $2.00 | $10.00 | $0.20 | $2.00 / $4.00 |
| kimi-k3.1-pro-highspeed | $4.00 | $20.00 | $0.20 | $4.00 / $8.00 |
| kimi-k3 (previous) | $3.00 | $15.00 | $0.30 | $3.00 / $6.00 |
The Highspeed build is the same weights served faster (officially 2-3x) at exactly double the price. The API is live today as kimi-k3.1-pro, alongside Kimi.ai, Kimi Work and Kimi Code; the reasoning tier is chosen with reasoning_effort (low / high / max, max by default).
Weights by Nov 4
API only for now — the weights are not out, and the official line is that full model weights will be released by Nov 4. The site logs it under the established convention:
| Model | ID | Status | Runs |
|---|---|---|---|
| Kimi-K3.1-Pro | 98 | RESTRICTED | zero |
| Kimi-K3.1-Pro-Highspeed | 99 | RESTRICTED | zero |
| Kimi-K3 (previous, 2.8T-A104B) | 73 | params backfilled | 56 |
We also backfilled two previously blank parameter fields: K3 (2.8T total / 104B active) and gemma-4-E4B-it (8B total / 4.5B effective).
The local-deployment reality: 2.8T parameters is terabytes even at NVFP4/MXFP4, a different order of magnitude from last batch's Flint (39.8B, GGUF Q4_K_XL 25.1GB) or Raon-1-mini (28.8B, Q4_K_M 19.7GB). The site's fit verdict for a 2.8T model is consistently "does not fit" — and the 56 existing K3 rows predate this release, so they say nothing about K3.1's speed. Whether the Nov 4 weight drop makes MXFP4 plus fast speculative decoding worth the trouble depends entirely on how thick the official quantized artifacts are.