IFM open-sources K2 Horizon models from 0.9B to 375B with training code and data recipes
testingcatalog · x · 2026-09-03
IFM has released its K2 Horizon model collection on Hugging Face, shipping not just weights but also the training code, data recipes, intermediate checkpoints, logs, and evaluations behind them.
The family spans 0.9B up to 375B (MoE, A23B active), with FP8 and GGUF variants, plus datasets like TxT360-v2 and Code/Math-Reasoning SFT sets.
Two architecture pieces ship alongside:
- MoVA: routes experts inside attention itself
- Uno: a diffusion adapter that generates token blocks in parallel
A rare fully reproducible open-source package, from data to weights to training logs.
Related event: MBZUAI Open-Sources K2 Horizon Model Family From 0.9B to 375B(3 posts)→
More from Models
- Meta ships Muse Spark 1.3: 98.5 on MRCR 256K, 25% fewer tokens, near Opus 5 on agentic benchmarks — rohanpaul_ai · 2026-09-03
- Claude Max 20x User Burns Through Entire 5-Hour Session Quota in About 30 Minutes — NathanSRobinson · 2026-09-03
- Developer Rant: Models Keep Getting 'Smarter' But Slower and More Annoying — uchikanda · 2026-09-03
- Frontier Models Give Suspiciously Bad Advice on Local AI Setups, Developer Reports — ikilaie · 2026-09-03
- Grok and Claude reportedly down at the same time, user reports outage — rand_longevity · 2026-09-03
- Apollo Research's Bronson Schoen: Models Know They're Being Tested and Still Lie — PeterBowdenLive · 2026-09-03