Moonshot releases open weights for Kimi K3.1 Pro
Moonshot released weights for its K3.1 Pro flagship, a 2.8T-parameter model with 104B active parameters and a 1M-token context.
Moonshot has released open weights for Kimi K3.1 Pro, its flagship model. It has 2.8T total parameters but only 104B active parameters. Modal describes it as a post-training update to K3, with reinforcement learning expanded for coding, financial research and long-horizon agentic work. The model accepts text and images and has a 1M-token context window.
The gap between total and active parameters is central to the model's design. It is a mixture-of-experts model, meaning it contains many specialist sub-networks and routes each token to only a few of them. A Reddit post on the release listed 896 experts, with 16 used per token. Total parameters set how much storage and memory the weights need. Active parameters set how much computation each token requires, and each token activates 104B parameters. Reddit commenters joked about running the model on consumer hardware, with one posting "My GTX 1030 is ready."
The independent evidence cited is Agent Arena. There, Kimi K3.1 Pro ranks first among open models and 13th overall, with a 5.1% net improvement across more than 17.5K agentic sessions. That is 4.2 percentage points better than K3. The model is also first among open models on Bash reliability and steerability. A single user reported that on an internal SRE-agent evaluation the model uses more tokens than alternatives but delivers "shockingly good" quality. That is one user's report.
Access comes through several routes. vLLM and SGLang shipped day-0 support, and the model is hosted on Ollama Cloud, Databricks and SkyPilot. On the license, Reddit commenters said personal use is permitted and most commercial use requires only attribution. They said stricter terms apply mainly to API providers.
What changed
Moonshot released open weights for Kimi K3.1 Pro, a model with 2.8T total parameters, 104B active parameters and a 1M-token context window.
Type: New capability · Applies to: Kimi K3.1 Pro
Why it matters
A model of this size with open weights gives companies a self-hostable option at frontier scale, though the hardware needed to run it is substantial.
What you can do
Teams evaluating open-weight coding and agent models should test Kimi K3.1 Pro on their own workloads, including token cost per task.
How the story developed
Adds Agent Arena's #1 open-model ranking, SGLang day-0 support, a 9.9% WeirdML score and an Artificial Analysis consistency test.
The commentary adds the presenter's reading of benchmark placements and the weight size and VRAM requirements.
The presenter adds the 96-part download size and the difficulty of running the model locally, and places it on several benchmarks.
Original source
We have no link to the original source for this story yet.
Further reporting
Where we saw it
Forward Future (Matthew Berman), Latent Space
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works