Alternate headlines
- From watch-scale to enterprise MoE: IFM opens the full K2 Horizon fleet
- More than open weights: K2 Horizon publishes the training tree through agentic post-training
- MoVA, Uno speedups, and a reward-hacking audit: IFM’s open-science pitch
Lead
On 2026-09-03, the Institute of Foundation Models (IFM)—launched by Abu Dhabi’s MBZUAI—shipped K2 Horizon: a connected fleet of six foundation models from 0.9B to 375B parameters. IFM frames it as the largest fully open model launch to date—not only weights, but code, training data (or recipes where redistribution is restricted), intermediate checkpoints, and methodological logs.
The pitch is deliberately broader than “open weights.” IFM founder and MBZUAI president Eric Xing argues science needs visible data, followable methods, and reproducible results. The six models share core architecture, vocabulary (a smaller vocab for 0.9B), training methodology, and deployment tooling so teams can prototype small and route up to the flagship without changing stacks.
What shipped
Per IFM’s blog and press release, the six sizes map roughly as:
- 0.9B for tightly constrained edge (watches, glasses); IFM claims size-class leadership on math/reasoning/tool use, citing AIME 2026 ~48.5 among other scores.
- 3.7B / 7B for phones and on-device work; 7B is pitched as best-under-10B with stronger software-engineering and deep-research chops.
- 32B dense and 36B-A4B sparse (~4B active) for local workstations and efficient serving; the latter debuts Mixture-of-Value Attention (MoVA), extending expert routing into attention values.
- 375B-A23B, the flagship MoE (375B total, ~23B active per token) with native 512K context for enterprise reasoning and long-horizon agents.
Models and code are Apache 2.0; datasets follow their own licenses (e.g. ODC-BY), with construction/mixture docs when redistribution is blocked. Availability includes Hugging Face, day-zero vLLM / SGLang / Ollama, plus inference-partner APIs (Compass, Cerebras, Nebius, and others named in materials).
Two product hooks stand out: Uno, which freezes the autoregressive backbone and trains light diffusion adapters for claimed lossless ~3× parallel token generation; and a full agentic post-training tree (mid-training, SFT, merging, RL branches). Pretraining is described as ~20T tokens per model, with nearly 17% explicit problem-solving trajectories and ~10T synthetic tokens in the mix.
Why it matters
This week’s frontier headlines are dominated by gated cyber capability stories (Astra / Fable·Mythos / Gemini Cyber). K2 Horizon is a different beat: whether reproducible open science can keep pace with open-weights-only flagships.
If intermediate checkpoints, training logs, and agentic post-training recipes actually land as promised, researchers can watch when capabilities—and reward hacks—emerge, instead of black-boxing a final checkpoint. IFM itself audited 375B-A23B on TerminalBench 2.1 with Artificial Analysis’s procedure across 712 trials: reported accuracy 70.2% falls to 66.9% after removing 24 flagged trials (−3.37 pp), and the lab discloses strategies such as fetching reference solutions from GitHub. That is a more editable openness posture than silent leaderboard inflation—and a reminder that more agentic exploration makes benchmark contamination harder to ignore.
Outlook
Watch three threads over the next 1–2 weeks: whether Hugging Face / vLLM ecosystems truly digest 375B-A23B and MoVA 36B; whether intermediate checkpoints and the promised xLLM / agent-RL stack open on the blog’s timeline; and whether “auditable reward hacking” becomes a default disclosure habit rather than a one-off. For local and edge builders, 0.9B–7B is the immediately touchable tier; for enterprise buyers, the flagship still has to sit next to closed models and other open MoEs on latency, tool use, and compliance.