Latest: Cerebras Runs Kimi K3.1 at 2,000+ Tokens/Sec on Wafer-Scale Hardware
The public weight release makes a large mixture-of-experts model available for self-hosted and commercial use under terms that target API resellers. The Cerebras demonstration shows the model can run at high inference speeds on specialized hardware.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
Cerebras released a video demonstrating its wafer-scale engine running Moonshot AI's Kimi K3.1 at over 2,000 tokens per second. The model whose weights it is running, Kimi K3.1 Pro, is a 2.8-trillion-parameter mixture-of-experts model Moonshot AI released publicly on Hugging Face on November 4, 2026. Only 104 billion of those parameters activate per token, and 69 of its 93 layers use linear attention to reduce the cost of its 1-million-token context window. The model ships at 4-bit precision, avoiding the 5.6 terabytes a 16-bit copy would require. Moonshot's modified MIT license exempts internal use but requires companies earning over $20 million annually from model hosting to negotiate a separate license.
Cerebras released a video on November 7 showing Kimi K3.1 running at over 2,000 tokens per second on its wafer-scale engine, the first reported third-party inference demonstration of Moonshot AI's open-weights model since its November 4 release.
Moonshot stated it intends to progressively raise the offensive cyber capability available in its open-weights releases over time, while confirming that Kimi K3.1 Pro's current level is intentionally throttled below what the company's pipeline can produce.
The 2.8-trillion-parameter model released November 4 also accepts image inputs alongside text, is available on HuggingChat, and is priced at $2/$10 per million tokens via API. Moonshot's own benchmarks claim it outperforms rivals roughly 8 to 15 times its size.
Moonshot AI published the weights of Kimi K3.1 Pro, a 2.8-trillion-parameter sparse model that activates only 104 billion parameters per token, on Hugging Face on November 4, 2026.
The release uses a modified MIT license requiring companies earning over $20 million annually from model hosting to negotiate a separate commercial license. Before releasing the weights, Moonshot said it removed cyber-exploit-related content from the model's mid-training corpus after detecting rising exploit skill, and that on Terminal-Bench 4.0 the model scores 65.7%, near Claude Opus 5.5's 66.4%.
Coverage describes Kimi K3.1 Pro's benchmarks, pricing, and intentionally throttled offensive cyber capability
BackgroundFrontier models, open models, and how they are compared
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.