Nov 05, 2026
Kimi K3.1 Pro: Specs, Price, Open Weights, and Benchmarks
Cost Optimization
Distributed Inference
Moonshot released Kimi K3.1 Pro on October 23 at $2 per million input tokens and $10 per million output, a third less than Kimi K3, and published the open weights on November 4 under the Kimi K3.1 Pro License. Specs, pricing, the model card, Moonshot's benchmark claims, and what it means for the hardware.

Moonshot's follow-up to Kimi K3 is out, and it costs less than the model it replaces. Here's what changed, what didn't, and what the open weights look like now that they're out.
Moonshot released Kimi K3.1 Pro on October 23, 2026, calling it the first model in the K3.1 series and its most capable open-weight model to date. It is a post-trained version of Kimi K3 with the same architecture and a much larger post-training budget, and it ships with a faster variant, Kimi K3.1 Pro Highspeed. Both are live on the Kimi API, and the full weights went up on Hugging Face on November 4 under the Kimi K3.1 Pro License.
Two things make this release different from a routine version bump. The price went down, not up: K3.1 Pro is a third cheaper than K3 on every rate. And because the architecture is unchanged, the self-hosting math for K3 applies to K3.1 Pro as written. This post covers the specs, the pricing, what Moonshot claims on benchmarks, and the route to using it now.
The short version
- Released: October 23, 2026, on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
- Architecture: same as Kimi K3. 2.8 trillion total parameters, 104B active per token, 1M-token context, text and image input.
- API price: $2.00 per million input tokens, $0.20 per million cached input, $10.00 per million output. Kimi K3 was $3.00 / $0.30 / $15.00.
- Highspeed variant: $4.00 / $0.20 / $20.00, which Moonshot says runs two to three times faster.
- Open weights: released November 4, 2026, on Hugging Face as moonshotai/Kimi-K3.1-Pro. 1.56 TB in MXFP4, under the Kimi K3.1 Pro License, a modified MIT.
- Benchmarks: Moonshot's own numbers put it within a point of Claude Opus 5.5 on Terminal-Bench 4.0 and ahead of GPT-6 Astra on most of the agentic rows.
The numbers that matter
| Spec | Kimi K3.1 Pro |
| Released | October 23, 2026 |
| Model IDs | kimi-k3.1-pro, kimi-k3.1-pro-highspeed |
| Total parameters | 2.8 trillion, sparse Mixture-of-Experts |
| Active parameters | 104B per token |
| Context window | 1M tokens |
| Inputs | Text and images |
| Relationship to K3 | Post-trained version of Kimi K3, same architecture |
| Open weights | Released November 4, 2026; moonshotai/Kimi-K3.1-Pro on Hugging Face, 1.56 TB, MXFP4 |
| License | Kimi K3.1 Pro License, a modified MIT; free use, with conditions for large hosted services and products over 100M monthly users |
| Access today | Kimi API, Kimi.com, Kimi Work, Kimi Code, and the weights on Hugging Face |
Moonshot describes the gains as coming from post-training: larger-scale reinforcement learning across hundreds of domains, with work on the RL framework, long-context inference, and sandbox infrastructure to make that scale possible. The base model is K3. That matters for anyone planning to run it, and we come back to it below.
Kimi K3.1 Pro pricing
Per million tokens, from Moonshot's API pricing page.
| Model | Input | Cached input | Output |
| Kimi K3.1 Pro | $2.00 | $0.20 | $10.00 |
| Kimi K3.1 Pro Highspeed | $4.00 | $0.20 | $20.00 |
| Kimi K3 | $3.00 | $0.30 | $15.00 |
Cache writes are billed separately on the K3-series models: $2.00 per million for a five-minute cache and $4.00 for a one-hour cache on K3.1 Pro, double that on Highspeed. Cached reads are billed only at the cache-hit rate. The Kimi K3.1 Pro pricing post has the full rate card, the caching math, and the Kimi Code plan prices.
The direction is the point. Moonshot cut the price of its flagship by a third while claiming a large capability jump, and the per-token cost of the new model now undercuts what K3 charged at launch in July. Highspeed costs exactly double for the same model served faster, which is the same trade Moonshot offers on its K2.7 Code line.
What Moonshot claims on benchmarks
All of these are Moonshot's own figures from its launch post, run with its own harness, and none has independent replication yet. The rows below are the ones most relevant to coding and agent work.
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra | DeepSeek V4.1 Flash | GLM 5.3 |
| Terminal-Bench 4.0 | 65.7 | 66.4 | 58.2 | 42.3 | 41.8 |
| DeepSWE v1.1 | 75.5 | 74.2 | 74.1 | 77.0 | 66.9 |
| SWE-Marathon v1.1 | 51.7 | 46.3 | 42.5 | 40.0 | 42.5 |
| HLE-Full, with tools | 70.0 | 67.7 | 57.2 | 58.8 | 62.5 |
| GPQA Diamond | 94.2 | 93.9 | 96.1 | 92.2 | 91.7 |
| Artificial Analysis Intelligence Index | 55 | 58 | Not reported | Not reported | 45 |
Read it with the usual caveats. Moonshot ran K3.1 Pro with the Claude Code harness and took the other models' scores from their own reports or public leaderboards, so the rows mix sources. The Intelligence Index is the one independent line, and there K3.1 Pro sits three points under Opus 5.5 and ten over GLM 5.3. Moonshot says that is an 11-point gain over Kimi K3 on the same index, and the jump within its own line is the clearest signal in the table.
Two other claims are worth noting. Moonshot says K3.1 Pro's claim-level hallucination rate on its internal test fell to 8.3% from 23.4% for K3. And it says it identified and removed cyber-related data from mid-training and the RL mix before the open release, trading some vulnerability-research capability for a safer open model. Both are Moonshot's numbers.
What it means for hardware
The model card settles it. K3.1 Pro ships as 96 safetensors shards totaling 1.56 TB, in MXFP4 weights with MXFP8 activations from quantization-aware training, the same format and the same size as K3. The serving requirement is unchanged: a multi-node cluster with 2 TB or more of aggregate GPU memory and fast interconnect. Moonshot lists vLLM, SGLang, and TokenSpeed as the recommended engines and states no minimum hardware. Everything in the Kimi K3 hardware requirements post applies to K3.1 Pro as written. The Kimi K3.1 Pro hardware requirements post has the GPU configurations and the license terms from the model card.
Which means the practical answer is the same one it was for K3. This is a model you call through an API unless you operate a cluster, and 104B active parameters is why the per-token price is what it is. The release of a Flash-class K3.1 would change that, and Moonshot hasn't announced one.
What a real workload costs
Take 10 million input tokens and 2 million output tokens a day, no caching, 30 days.
| Model and route | Input / output per 1M | Per day | Per month |
| DeepSeek V4 Pro (Yotta AI Gateway) | $0.99 / $2.97 | $15.84 | $475.20 |
| GLM 5.3 (Yotta AI Gateway) | $1.40 / $4.40 | $22.80 | $684.00 |
| Kimi K3.1 Pro (Kimi API) | $2.00 / $10.00 | $40.00 | $1,200.00 |
| Kimi K3 (Yotta AI Gateway) | $3.00 / $15.00 | $60.00 | $1,800.00 |
| Muse Wonder (Meta API) | $4.00 / $20.00 | $80.00 | $2,400.00 |
| Kimi K3.1 Pro Highspeed (Kimi API) | $4.00 / $20.00 | $80.00 | $2,400.00 |
| GPT-6 Astra (OpenAI API) | $10.00 / $50.00 | $200.00 | $6,000.00 |
K3.1 Pro lands at half of Muse Wonder's bill and a fifth of GPT-6 Astra's, for a model Moonshot puts within a point of Opus 5.5 on the main coding-agent benchmark. It is still two and a half times the cost of DeepSeek V4 Pro and nearly double GLM 5.3, which is the usual shape of this market: the frontier-adjacent open model costs more per token than the open flagships a tier below it, and the question is whether the capability difference shows up in your output.
How to use it today
Call it on Moonshot's API with the model ID kimi-k3.1-pro, or kimi-k3.1-pro-highspeed if latency matters more than cost. The API is OpenAI-compatible, so an existing client switches with a base URL and a model string.
For the models it competes with, Yotta AI Gateway puts Kimi K3, DeepSeek V4 Pro, GLM 5.3, and Qwen3.8-Max behind one OpenAI-compatible key with flat per-token pricing, so testing K3.1 Pro against them is a model-string change in the same client. The best Chinese LLMs roundup covers how those four labs' models split up by job.
The weights are on Hugging Face as moonshotai/Kimi-K3.1-Pro, so the self-hosting route is open for teams with the cluster to run it. The license is a modified MIT: free to use, modify, and redistribute, with two conditions. Running it as a hosted model service that brings in more than $20 million in any 12 months requires a separate agreement with Moonshot, and products with more than 100 million monthly active users or $20 million in monthly revenue have to display Kimi K3.1 Pro in the interface. Internal use is exempt from both.
Frequently asked questions
What is Kimi K3.1 Pro? Moonshot's new flagship, released October 23, 2026. It is a post-trained version of Kimi K3 with the same 2.8T total / 104B active architecture, a 1M-token context, and text and image input.
How much does Kimi K3.1 Pro cost? $2.00 per million input tokens, $0.20 per million cached input, and $10.00 per million output on Moonshot's API. The Highspeed variant is $4.00 / $0.20 / $20.00.
Is Kimi K3.1 Pro cheaper than Kimi K3? Yes, by a third on every rate. K3 is $3.00 / $0.30 / $15.00.
Is Kimi K3.1 Pro open source? The weights are public on Hugging Face under the Kimi K3.1 Pro License, a modified MIT with a revenue condition for large hosted services and an attribution condition for very large products. That is open weight, not an OSI-approved open-source license, the same approach Moonshot took with Kimi K3.
What is the difference between K3.1 Pro and K3.1 Pro Highspeed? Same model, served two to three times faster according to Moonshot, at double the price.
How does Kimi K3.1 Pro compare to Claude Opus 5.5? On Moonshot's own numbers, within a point on Terminal-Bench 4.0 (65.7 against 66.4), ahead on DeepSWE v1.1 and SWE-Marathon, and three points behind on the Artificial Analysis Intelligence Index. These are vendor-reported and not yet replicated.
How much VRAM does Kimi K3.1 Pro need? The same as Kimi K3: the checkpoint is 1.56 TB in MXFP4, so plan on a multi-node cluster with 2 TB or more of GPU memory. See the Kimi K3 hardware requirements post.
Is Kimi K3.1 Pro on Yotta? Kimi K3 is live on Yotta AI Gateway at $3.00 per million input and $15.00 per million output, alongside DeepSeek V4 Pro, GLM 5.3, and Qwen3.8-Max.
Are the Kimi K3.1 Pro weights released? Yes. Moonshot published them on Hugging Face on November 4, 2026, as moonshotai/Kimi-K3.1-Pro: 2.8T parameters, MXFP4, 1.56 TB, Kimi K3.1 Pro License.
Bottom line
Kimi K3.1 Pro is the same 2.8T model as K3 with far more post-training, priced a third lower, and Moonshot is claiming frontier-level results on coding-agent benchmarks that no one has independently checked yet. The weights are out at the same 1.56 TB as K3, so the hardware story is K3's hardware story.
For most teams the decision is the familiar one. Call it through an API and measure it against the open flagships you already run. On Yotta AI Gateway that comparison set is one key away, and the GPT-6 Astra pricing and Muse Wonder pricing posts cover the closed-model side of the same bill.



