Skip to content
Main Site News Console

[AINews] Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-Class Open Weights Narrow Closed-Source AI Lead to 1 Month... Plus 2–3x Faster Pro High-Speed API

· Latent Space Translated
播客深度访谈

Two months ago, we reported on the prediction by Z.ai CEO Prof. Jie Tang that we would see Fable-level open-weights models before the end of the year. Two weeks ago, Meta hit Fable-level among closed-source models, but with 2.5 months left until year-end, we finally have it—a 2.8T (2.8 trillion parameter) open-weights model that surpasses Fable 5.1, not just Fable 5:

View post on X ↗

And with higher efficiency:

View post on X ↗

You can tell how proud Kimi is of their latest model just from how beautifully their product landing page is designed:

As usual, we didn’t get many research details—this update is somewhat similar to Z.ai’s 5.2 to 5.3 bump—a minor version bump backed by a massive influx of post-training compute:

Kimi K3.1 Pro is a post-trained version of Kimi K3, backed by significantly more post-training compute. To enable large-scale joint reinforcement learning (joint RL) across hundreds of downstream domains, we optimized every layer of the technical stack—including the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for environment quality checks—ensuring that every incremental unit of RL compute delivers reliable, predictable gains. As a result, K3.1 Pro achieves consistent improvements across every domain.


AI Twitter Recap

Top News: Kimi K3.1 Pro Launch and Reactions

Highlights

Moonshot has launched Kimi K3.1 Pro, delivering near-frontier performance across independent third-party benchmarks with a dramatically scaled-up model architecture, and committed to open-sourcing the weights on November 4, 2026.

  • Model Details: Moonshot describes K3.1 Pro as a multimodal, “3T-class” Mixture-of-Experts (MoE) model featuring 2.8 trillion total parameters, 104 billion active parameters, and a 1-million-token context window. Its claimed strengths include full-stack software engineering, Blender/CAD modeling, financial modeling, and long-horizon autonomous agents. @Kimi_Moonshot

  • Availability: Both the standard and Highspeed editions are now fully live across Kimi’s web/app, Kimi Work, Kimi Code, and the API. Moonshot claims the Highspeed edition delivers 2–3x faster generation speeds while maintaining the same intelligence level; the official summary notes it is priced at twice the standard rate. The release also introduced Kimi Hosted Agents, a managed runtime framework offering a 24/7 execution environment built specifically for long-horizon tasks and multi-agent orchestration. @ZhihuFrontier, @ZhihuFrontier

  • Independent General Benchmark Results: Artificial Analysis ranked K3.1 Pro 4th on its Intelligence Index with a score of 55, trailing only Opus 5.5 (58) and Sonnet 5.5 and Muse Wonder Full (both at 56). Its most notable gains were concentrated in agentic capabilities rather than being evenly distributed across knowledge and science evaluations. @ArtificialAnlys

  • Independent Coding Benchmark Results: Arena placed it 8th on Code Arena: WebDev with a score of 1,692, climbing 9 ranks and 37 points compared to Kimi K3. This pushed Moonshot back into Arena’s top four labs. @arena

  • Independent Cybersecurity Benchmark Results: Despite intentionally pruning cybersecurity-related training data, it went on to score 64% on the Artificial Analysis Cyber Index—the highest score among all publicly accessible models, ranking second overall behind only the trusted-access-tier GPT-6 Sol Daybreak Blue (69%). @ArtificialAnlys

#播客#深度访谈#API#4Allapi.com

Published by the 4All API team

Original link:https://latent.space/p/ainews-kimi-31/

Need a mainstream LLM API? 4All API gives you one key to call OpenAI, Anthropic, Google Gemini, Qwen, DeepSeek, and dozens more — at official-pass-through pricing, with enterprise-grade reliability, integrated in 5 minutes.

Sign up for the 4All API console →