Much awaited Kimi-K2 Technical Report is Out Now!. Kimi K2 is 1T-parameter open-weight MoE model built for agentic intelligence. Using MuonClip optimizer and a 15.5T-token high-quality dataset, Kimi K2 achieves stable, scalable pre-training. Post-training combines large-scale

RT @Kimi_Moonshot: We're building the Kimi Discord community! . If you're passionate about AI, love Kimi, or want to help shape the future….

RT @Teknium1: Little do people know but interstellar, on work for the Hermes 3 function calling capabilities - built an almost identical pi….

read here:.

Mixture of Raytraced Experts. —. - MRE replaces fixed top-k MoE gating with a dynamic, stochastic raytracing mechanism. - firing ray probabilistically activates a sequence of experts using a routing net, like a poisson walk on a softmax graph. - no load balancing, hard top-k,

Nous released their Hermes 3 dataset. --.- 1m samples.- uncensored sota for its time across llama-3 (8b, 70b, 405b).- dense in-prompt adherence, roleplay, subjective/objective tasks.- rich tool use, structured outputs, api-like call patterns.- early agentic traces: xml-tagged

RT @teortaxesTex: Oh.

The 1 Trillion param Open-Source Agentic Model from Kimi Moonshot is live. Kimi K-2 ✌🏻. (based on MuonClip Optimizer). get started here:

RT @teortaxesTex: Claim. Source: “revealed in a dream”. I do sometimes get leaks in dreams, and Arjun is Indian so maybe it's a Ramanujan s….

Kimi’s next drop gonna be spicy 🥵

SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning.via Multi-Agent Multi-Turn Reinforcement Learning. —.adversarial selfplay can yield cognitive dividends in LLMs trained for general reasoning without explicit labels or human reward shaping. instead of finetuning on

Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search. —.AB‑MCTS (Adaptive Branching Monte Carlo Tree Search) inference-time LLMs strategy. repeated sampling w iterative refinement, guided by external feedback. each node uses bayesian posterior

Ovis-U1, a 3B parameter unified MLLM. built on a RoPE-based diffusion visual decoder and a CLS-token-based refiner. integrates 3 core capabilities .- image comprehension.- text-to-image synthesis.- and fine-grained editing.all within a single architecture. performance wise, .it

if money could buy good AI.siri would have been the best model. a lesson in there.