Evidence at a glance
The mechanism in one line
Compress the visual or contextual input before the main reasoning path.
Route or verify the expensive step instead of repeating the full path.
Translate the mechanism into a bounded deployment or evaluation check.
“Open Weight” Is Not Yet a Deployment Option
Moonshot AI introduced Kimi K3.1 Pro in October 2026 as a post-trained version of Kimi K3, aimed at multi-step tasks such as coding, 3D creation, and financial research. It is not presented as a new architecture. Instead, it adds post-training compute to the existing K3 foundation and is currently offered through an API and Kimi products.
That distinction defines what technical leads can evaluate today. K3.1 Pro has a one-million-token context window. Third-party materials list 2.8 trillion total parameters and 104 billion active parameters per token, but the full weights are not yet available. Moonshot has promised to release them by November 4, 2026. Until then, teams can test the hosted service’s capability, latency, and bill, but not its self-hosted memory requirements, throughput, or operational costs.
The Change Is in Training, Not the Parameter Count
Moonshot says K3.1 Pro uses substantially more post-training compute and joint reinforcement learning across hundreds of downstream domains. To support that work, the team also optimized its reinforcement-learning framework, long-context inference, sandbox infrastructure, rollout scheduling, and environment quality checks. The intended upgrade is therefore better action over multi-step tasks and tool environments, not simply a larger model.
The underlying design carries over from K3: Kimi Delta Attention and Attention Residuals, alongside a Stable LatentMoE mixture-of-experts framework. The source material says the model activates 16 of 896 experts, but sparse activation should not be mistaken for single-GPU deployability. Moonshot’s deployment guidance for K3 recommends supernodes with 64 or more accelerators. K3.1 Pro’s open-weight deployment requirements remain to be confirmed after release.
The Scores Show Both Strengths and Limits
In Moonshot’s published coding results, K3.1 Pro scores 65.7% on Terminal-Bench 4.0, just below Claude Opus 5.5 at 66.4% in the same table. On SWE-Marathon, it scores 51.7%, ahead of Opus 5.5 at 46.3%. These figures are not a universal leaderboard: they are self-reported by Moonshot, and the Terminal-Bench result uses the Claude Code harness at maximum reasoning effort. Artificial Analysis reports 67.2% using the mini-swe-agent harness. Different harnesses should not be treated as one measurement scale.
Performance is uneven across task types. On Moonshot’s table, K3.1 Pro scores 38.1% on Terminal-Bench Science 0.1, compared with 68.1% for GPT-6 Astra and 63.3% for Opus 5.5. On tool-assisted BenchCAD, it scores 92.7%, below the other two models’ 96.2% and 95.9%. Its score on Kimi’s internal long-horizon coding benchmark rises from K3’s 55.95 to 61.2, but that is an internal test. For engineering teams, these contrasts are more useful than a broad claim of frontier performance: test against your own task mix before assigning it scientific reasoning, CAD, or long-running coding work.
A Lower Price Does Not Guarantee a Lower Total Cost
K3.1 Pro’s API list price is $2 per million input tokens and $10 per million output tokens, with cache hits priced at $0.20. The previous K3 price was $3 for input and $15 for output. A Highspeed variant is priced at $4 and $20 respectively, and Moonshot says it runs two to three times faster. The lower base price may matter for high-volume workloads, but the Highspeed premium and the reasoning effort a task requires also shape the total bill.
Thinking mode is always on, with low, high, and maximum reasoning effort available. The API is OpenAI-compatible, and the model is accessible through the Kimi API and products such as Kimi Code. A one-million-token context is a capacity limit, not a promise that every request of that length will be inexpensive. A useful evaluation should track task success, retries, input and output tokens, latency, and human handoffs—not just the listed price per million tokens.
Risk Reduction Is Not a Substitute for Governance
Moonshot says it removed cybersecurity-related data or tasks from mid-training data and reinforcement-learning environments for the open-weight release, while retaining general programming tasks. On its internal xdev-bench-hard test, the unfiltered checkpoint scores 90.5% pass@3, compared with 81.0% for the released version. This suggests that targeted filtering may reduce capability on particular offensive tasks, but the result is an internal company evaluation, not an independent replication or proof of reliability across all safety scenarios.
Moonshot also reports that claim-level hallucinations on its internal Kimi Hallu Bench Hard fell from 23.4% for K3 to 8.3% for K3.1 Pro; GPT-6 Astra scores 5.5% on the same test. These results are useful prompts for further testing, not substitutes for validation in a team’s own domain. The practical judgment should remain layered: try the API and check benchmark conditions now; do not describe the model as self-hostable before its weights arrive; and, after release, independently assess deployment costs, licensing, and safety behavior.