Kimi K3.1 Pro
Moonshot's flagship open-weight MoE giant — 2.8T total parameters, 1M-token context, vision, tool calling and top-tier agentic coding.
Overview
What is Kimi K3.1 Pro
Kimi K3.1 Pro is Moonshot AI's flagship open-weight mixture-of-experts model — 2.8T total parameters with 104B active — released in October 2026 with a 1M-token context window. It combines vision input, tool calling, web search and adjustable reasoning with leading agentic-coding performance, and runs on Venice with zero retention of your prompts.
Running it privately on Venice
On Venice, Kimi K3.1 Pro runs under the private tier with zero retention — your prompts, uploads and outputs are never stored, profiled, or used for training, and you don't need a permissioned enterprise agreement to use a frontier-class model. Note the honest boundaries: inference on Venice for this model is not TEE-isolated and not end-to-end encrypted — the privacy guarantee is the zero-retention policy itself. You also get per-token billing instead of a subscription, so occasional heavy-context sessions stay pay-as-you-go.
Assessment
Strengths and limitations
- Top-tier agentic coding: Moonshot reports 65.7% on Terminal-Bench 4.0, and Artificial Analysis measured the highest Terminal-Bench 4.0 score of any model at launch.
- A full 1M-token context window with sparse-attention long-context modeling, plus cached input at $0.20 per 1M tokens on Venice — a 90% discount for repeated long-document work.
- Genuinely multimodal: accepts multiple images per request alongside text, with vision, function calling, web search, JSON-schema structured output and adjustable reasoning effort all exposed through the API.
- Open weights under a permissive custom licence: you can self-host or fine-tune it, which almost no model at this intelligence level allows.
- Large agentic gains over its predecessor: AutomationBench-AA rose from 58% to 73% and GDPval-AA Elo rose 312 points over Kimi K3, per Artificial Analysis.
- Slow: Artificial Analysis measured a 146-second time to first token, ~32 tokens/second output, and roughly 37 minutes per Intelligence Index task — plan for long waits on big jobs.
- Expensive for an open-weights model: $2 in / $10 out per 1M tokens is several times the median, and it is verbose (~55K output tokens per index task).
- Independent coding evals are more sobering than launch numbers: one third-party benchmark scored it 72% on agentic coding, behind Claude Sonnet 5.5 (75%) at the same per-token price.
- Not uncensored: Venice classifies this model as filtered, so it is not the pick for unfiltered output.
- On Venice it runs with zero retention but without TEE isolation or end-to-end encryption.
Capabilities
What it supports
- Tool use / function calling
- Vision — image input
- Multiple image inputs per request
- Adjustable reasoning effort
- Web search grounding
- Structured output via JSON schema
- Code-optimized tuning for engineering tasks
- Self-hosting from released weights
- Uncensored / unfiltered mode
- TEE-isolated inference on Venice
- End-to-end encrypted prompting
Specifications
Datasheet
- Maker
- Moonshot AI
- Modality
- Text and image input, text output
- License
- Custom Moonshot AI licence — free to use, modify, self-host and redistribute; MaaS providers with over $20M annual revenue need a separate agreement
- Context window
- 1,000K tokens
- Released
- October 23, 2026
- Architecture
- Mixture-of-Experts (MoE), 93 layers, 896 experts with 16 selected per token plus 2 shared experts
- Parameters
- 2.8T total, 104B activated per token
- Attention mechanism
- KDA and Gated DeepSeek Sparse Attention (Gated DSA) for efficient long-context modeling
- Vision encoder
- MoonViT-V2, 401M parameters
- Native quantization
- MXFP4 weights / MXFP8 activations (quantization-aware training)
- Max output
- 128K tokens
- Capabilities
- Vision, Function calling, Reasoning, Web search, Code-optimized
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Oct 2026
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "kimi-k3-1-pro",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $2 per 1M input tokens and $10 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Best for | Context | Open weights | Price (Venice) |
|---|---|---|---|---|
| Kimi K3.1 Pro | Agentic coding & long-context research | 1M tokens | No | $2 in · $10 out / 1M |
| Qwen 3.8 Max | General reasoning at scale | 1M tokens | No | $2.50 in · $7.50 out / 1M |
| Gemini 3.8 Flash | Speed & cost efficiency | 1M tokens | No | $0.94 in · $4.69 out / 1M |
| Muse Wonder | Peak reasoning quality | 1M tokens | No | $5 in · $25 out / 1M |
| Grok 4.7 | Realtime & general assistant work | 500K tokens | No | $2.27 in · $6.80 out / 1M |
Pick Kimi K3.1 Pro when you need the strongest open-weights intelligence for agentic coding or million-token research and can tolerate slow, expensive inference; take Gemini 3.8 Flash for speed on a budget or Muse Wonder when only the absolute intelligence ceiling matters.
Use cases
What it is good for
- 01Long-horizon agentic coding: multi-step terminal work, refactors and tool-driven engineering sessions across a 1M-token codebase context.
- 02Financial and legal research over very long documents, with cached input keeping repeated queries on the same corpus cheap.
- 03Multimodal analysis combining screenshots, diagrams and text in a single request.
- 04Structured data extraction pipelines using JSON-schema outputs and function calling.
- 05Self-hosted deployments where teams need frontier-class intelligence but cannot ship prompts to a closed vendor.
Prompting
Getting better results
Set reasoning effort deliberately — Moonshot's headline scores come at max effort; lower it for quick queries to cut latency and token spend.
Reuse long context across turns so input hits the cached tier at $0.20 per 1M tokens instead of the full $2.
Attach multiple images (screenshots, charts, mockups) in one request rather than spreading them across turns — the model accepts several per call.
Pass a JSON schema when you need machine-readable output instead of parsing free-form prose.
For coding tasks, let it plan and call tools in a loop rather than asking for a one-shot answer — its strength is multi-step execution.
Enable web search on factual questions; its hallucination rate improves markedly when grounded.
Version history
Predecessor — 44 on the Artificial Analysis Intelligence Index; K3.1 Pro gains 11 points.
Current flagship — 2.8T MoE, 1M context, vision, agentic focus.
Same model family tuned for 2–3× faster responses.
FAQ
Frequently asked questions
Kimi K3.1 Pro is Moonshot AI's flagship open-weight mixture-of-experts model, released in October 2026 as the first model in the K3.1 series. It has 2.8T total parameters (104B active per token), a 1M-token context window, native vision input, and is post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of domains.
Venice bills Kimi K3.1 Pro per token: $2 per 1M input tokens and $10 per 1M output tokens, with cached input at just $0.20 per 1M — a 90% discount that makes repeated work over the same long documents dramatically cheaper. There is no subscription required.
The weights are open — Moonshot publishes them under a custom licence that permits free use, modification, self-hosting and redistribution. It is not OSI-style open source in the purest sense: if you run it as a Model-as-a-Service business with over $20M in annual revenue, the licence requires a separate agreement with Moonshot AI.
Yes. Kimi K3.1 Pro supports function calling, web search grounding, JSON-schema structured output, and adjustable reasoning effort, alongside vision input with multiple images per request. It is explicitly code-optimized, which is where its agentic benchmark results concentrate.
Yes. It is natively multimodal with a 401M-parameter MoonViT-V2 vision encoder, and it accepts multiple images in a single request — useful for comparing screenshots, charts or documents side by side within the same 1M-token context.
Both offer a 1M-token context on Venice. Kimi K3.1 Pro is the open-weights option with stronger agentic-coding results and cheaper output per token ($10 vs $7.50 is actually in Qwen's favor on output, but K3.1 Pro wins on input at $2 vs $2.50 and adds self-hosting). Choose K3.1 Pro for agentic engineering and open weights; Qwen 3.8 Max if you want a closed, general-purpose alternative at similar pricing.
Not fast. Independent measurement by Artificial Analysis recorded roughly 146 seconds to the first output token, about 32 tokens per second, and around 37 minutes per Intelligence Index task — the model reasons at length (about 55K output tokens per task). Budget time accordingly, or lower the reasoning effort for lighter queries.
Yes — Venice runs it under the private tier with zero retention: prompts, images and outputs are not stored, profiled, or used for training, and no personal history is attached to your usage. Be aware that this model's inference is not TEE-isolated or end-to-end encrypted; the guarantee is the zero-retention policy itself.
Moonshot AI released Kimi K3.1 Pro on October 23, 2026, alongside a faster K3.1 Pro Highspeed variant, with the full model weights published shortly after. It has been available on Venice since October 2026.
Run Kimi K3.1 Pro privately
No prompt logging. No data used for training.