Meta gives you Llama’s weights. Alibaba gives you Qwen’s weights. On September 3, MBZUAI’s Institute of Foundation Models released K2 Horizon — six models from 0.9B to 375B — and opened the training data, training code, and methodology too. All Apache 2.0. The 375B version is the largest fully open model ever shipped. HackerNews put it on the front page at 271 points.
A fleet, not a flagship
Each size has a job: 0.9B for local dev, 7B for cheap deployment, 32B for daily heavy use, 375B (a MoE with 23B active parameters, 524k context) for long-horizon agents. The small ones set new records on math, reasoning, and coding benchmarks in their size class. The 375B scores 47 on Artificial Analysis’s Intelligence Index — 11th of 112 comparable models, where the median is 29.
The technical hook: diffusion distillation generates token blocks in parallel — roughly 3x faster inference, no quality loss.
How you use it
No API gatekeeping. Download the weights, self-host, fine-tune, or reproduce training from scratch with the published data and code. For teams whose compliance won’t touch models with opaque training data, this is the one open model that answers every question.
You Might Also Like
- Code Arena Finally Gives Developers a Fair way to Judge ai Coding Models
- From Claude Flow to Ruflo 22k Stars 5900 Commits and the Multi Agent Swarm Taking Over Claude Code
- Oh my Claudecode Turns Claude Code Into a 32 Agent dev Team and 11 8k Stars Agree
- Pi Mono 29k Stars and a 200 Token System Prompt That Rivals Claude Code
- Claude Code Source Leak Kairos how a 59 8 mb File Exposed Anthropics Entire Agent Playbook
Top AI Product 
Leave a comment