Microsoft Foundry Adds Moonshot AI's Kimi K3.1 Pro With 2.8T Parameters

Microsoft has added Moonshot AI’s Kimi K3.1 Pro to its model catalog, according to the Foundry announcement . The managed listing gives developers an API path to a 2.8-trillion-parameter open-weight model with text and image input and a context window of up to 1,048,576 tokens.

Moonshot designed K3.1 Pro for software agents that plan multistep tasks, call tools, inspect results, and revise their work. Teams already using Foundry can now evaluate it alongside models from OpenAI, DeepSeek, Mistral, and Qwen through familiar deployment, billing, and governance workflows.

2.8T parameters, 104B active

K3.1 Pro uses a mixture-of-experts architecture. The network contains 2.8 trillion parameters, but it routes each token through 104 billion of them. Activating a subset lowers the computation required for each token while preserving the capacity of a much larger model.

Moonshot retained the earlier K3 architecture and updated its post-training. The release uses MXFP4 quantization-aware training, which prepares the model to operate with a compact 4-bit numeric format. Its 401-million-parameter MoonViT-V2 vision encoder converts images into representations the language model can process.

The 1,048,576-token context window can hold prompts, retrieved files, tool traces, image-derived tokens, and generated output, subject to the limits of the deployed endpoint. That capacity targets long codebases and extended agent sessions, although larger contexts also increase latency and input costs.

Moonshot published the weights on Hugging Face on November 4, 2026, under the custom Kimi K3.1 Pro License. The 1.56 TB download requires substantial storage, accelerator capacity, and serving infrastructure. Open weights make the parameters available for download; usage rights remain governed by Moonshot’s license.

Benchmark leads, with a harness caveat

Moonshot reports strong results on coding and agent evaluations, including a narrow gap behind Claude Opus 5.5 on Terminal-Bench 4.0 and the highest listed score on SWE Marathon v1.1.

Vals AI ranks the model ninth among 34 models on the Vals Index with a score of 62.52%. Its strongest listed result is third among 61 models on Terminal-Bench 4.0.

The comparisons are not fully controlled. Moonshot evaluated K3.1 Pro with the Claude Code harness while taking competing scores from model reports or public leaderboards. Differences in harnesses, prompts, tool configurations, and retry policies can materially affect agent benchmarks, so the table supports evaluation rather than a definitive model ranking.

Moonshot also reports lower factual error rates on its internal Kimi Hallu Bench Hard test. The claim-level hallucination rate fell from 23.4% for K3 to 8.3% for K3.1 Pro. At the response level, the rate of answers containing at least one factual error fell to 42.4% of K3’s rate. These figures come from an internal benchmark and require independent validation.

Foundry prices output at $11.50 per million

Microsoft lists one hosted price tier for K3.1 Pro, with separate rates for uncached input, cached input, and generated output.

At those rates, a request with 100,000 uncached input tokens and 10,000 output tokens costs $0.345. The same request costs $0.138 when all input tokens qualify for the cached rate. Agent evaluations should track total spending per completed task because retries, tool loops, and long histories can outweigh the advertised per-token price.

Moonshot removed cyber training data

Moonshot says it removed cyber-related data from both mid-training and the reinforcement-learning post-training mix before releasing the weights. The change reduced performance on the company’s internal xdev-bench-hard security evaluation.

A checkpoint trained without those controls solved 90.5% of tasks at pass@3, meaning at least one of three attempts succeeded. The released K3.1 Pro scored 81.0%, a reduction of 9.5 percentage points. The difference is most relevant to vulnerability research, exploit analysis, and autonomous security workflows.

Foundry handles the deployment layer

A managed Foundry endpoint removes the need to download and serve 1.56 TB of weights. Organizations already using the platform may also be able to reuse existing procurement, identity, networking, monitoring, and billing controls.

The catalog listing still leaves deployment-specific questions for each organization. Teams need to confirm regional availability, quotas, data retention terms, network options, endpoint context limits, model-version policies, and the permissions granted by Moonshot’s custom license.

Build the evaluation around completed tasks

• Task quality: Compare successful completion rates on representative coding, research, and document workflows.
• Tool use: Test function selection, schema adherence, error recovery, retries, and behavior after failed calls.
• Long context: Measure retrieval accuracy, latency, and cost as prompts approach production sizes.
• Vision: Evaluate the diagrams, screenshots, charts, and scanned documents used by the application.
• Factuality: Check claims against trusted sources and measure both claim-level and response-level errors.
• Security work: Test workflows affected by the deliberate reduction in cyber capability.
• Operations: Verify throughput, rate limits, regional support, logging, data controls, and license terms.

K3.1 Pro combines a million-token context window, multimodal input, competitive agent benchmarks, and managed Foundry access. Its mixed-source benchmark comparisons and custom license warrant scrutiny, while the hosted endpoint gives developers a practical way to measure its quality, latency, and cost against models already in production.