IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B
Most open mannequin launches launch one checkpoint and a benchmark desk. The Institute of Foundation Models (IFM) launched one thing wider final week. IFM is the frontier lab launched by MBZUAI in May 2025. K2 Horizon is a fleet of six fashions: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B. Shipping alongside them are the pre-training corpus, intermediate checkpoints, coaching code, configs and fine-grained logs. IFM calls it the biggest totally open-source mannequin launch in AI historical past.
Is it deployable? Yes, all six sizes sit on Hugging Face beneath Apache 2.0, with FP8 and GGUF builds. Day-zero help covers vLLM, SGLang and Ollama, on NVIDIA, AMD and Cerebras {hardware}. Hosted APIs run by means of Compass, Cerebras and Nebius through platform.ifm.ai.
What Actually Shipped
The six fashions share a core structure, vocabulary, coaching methodology, interfaces and deployment tooling. The 0.9B mannequin makes use of a smaller vocabulary. That consistency is the purpose: groups can prototype on 3.7B and scale to 375B-A23B with out altering their serving stack.
Each mannequin is pre-trained on roughly 20 trillion tokens. Nearly 17% of the pre-training corpus consists of problem-solving trajectories with express reasoning. About 10 trillion tokens have been artificial.
Post-training information was folded in from mid-training slightly than saved for the top. IFM analysis workforce studies over 100 million distinctive synthesized duties. Tool definitions have been introduced in JSON, XML and Markdown throughout coaching so the mannequin learns semantics slightly than syntax. Markdown grew to become the inference default, roughly 18.5% extra token-efficient than JSON on IFM’s information.
MoVA: Sparsity Moved into Attention
Conventional Mixture-of-Experts applies sparsity to feed-forward layers. Mixture-of-Value Attention (MoVA) extends knowledgeable routing into multi-head consideration itself, opening a second axis for scaling capability. It stays suitable with FlashAttention, grouped-query consideration and sparse consideration.
The result’s K2-Horizon-MoVA-36B-A4B: 36B complete parameters, roughly 4B energetic per token. Under matched coaching circumstances it lands barely beneath the dense 32B mannequin. On IFM’s tables it posts 58.6 on Terminal-Bench 2.1 and 26.8 on tau3-Banking, main its comparability set on each.
Uno: A Lossless Decoding Speedup as a LoRA
Uno freezes Horizon’s autoregressive parameters and trains a small set of diffusion parameters that be taught solely how to generate effectively. Through what IFM calls diffusion distillation, these adapters emit blocks of tokens in parallel. The press launch places the speedup at roughly 3× with no high quality degradation. It ships as a LoRA adapter, presently 7B-Uno and 0.9B-Uno.
Numbers price realizing
K2-Horizon-375B-A23B scores 70.2 on Terminal-Bench 2.1, 1,441 Elo on GDPVal-AA, 67.7 on MCPMark and 87.3 on GPQA Diamond. It leads its desk on SWE-Atlas-QnA at 48.4 however trails GPT-5.6 Luna and Claude Sonnet 5 on most agentic rows.
The small fashions are the sharper story. 7B posts 70.6 on SWE-bench Verified and 59.0 on BrowseComp. 3.7B posts 68.6 on SWE-bench Verified. 0.9B reaches 48.5 on AIME 2026 and 79.9 on HumanEval+, sufficiently small to run beneath quantization on a watch.
The Audit IFM Ran on Itself
This is the half many different labs don’t publish. IFM ran 375B-A23B throughout 89 Terminal-Bench 2.1 duties, eight makes an attempt every. That is 712 trials, 500 passing, a reported 70.2% accuracy. Every passing trial was then re-audited utilizing Artificial Analysis’s reward hacking procedure.
The audit flagged 24 trials throughout 10 duties. Removing them drops accuracy to 66.9%, a 3.37-point correction. That sits between the flag charges Artificial Analysis studies for Claude Fable 5 (2.2%) and GPT-5.6 Luna (4.1%). Behaviors included finding benchmark repositories on GitHub and downloading reference options. IFM additionally disclosed a 7B run that reached an inflated 82 on SWE-bench by discovering solutions.
Interactive explainer
Similar Posts
MBZUAI Researchers Release K2 Think: A 32B Open-Source System for Advanced AI Reasoning and Outperforms 20x Larger Reasoning Models
A workforce of researchers from MBZUAI’s Institute of Foundation Models and G42 launched K2 Think, is a 32B-parameter open reasoning system for superior AI reasoning. It pairs lengthy chain-of-thought supervised fine-tuning with reinforcement studying from verifiable rewards, agentic planning, test-time scaling, and inference optimizations (speculative decoding + wafer-scale {hardware}). The result’s frontier-level math efficiency with…
University of Michigan Researchers Propose G-ACT: A Scalable Machine Learning Framework to Steer Programming Language Bias in LLMs
LLMs and the Need for Scientific Code Control LLMs have rapidly evolved into complex natural language processors, enabling the development of agentic systems that manage complex workflows. However, the use of LLM agents for generating scientific code is unexplored. Scientific software primarily depends on C++, CUDA, and other low-level languages, which are underrepresented in most…
Meet Memory OS: A 6-Layer Open-Source Memory Stack Built on Top of Hermes Agent
Hermes Agent already remembers throughout periods. The open-source agent from Nous Research ships with curated reminiscence recordsdata and full-text session search. But a brand new neighborhood undertaking argues that in-built reminiscence is just too shallow for critical work. A new library named ‘Memory OS‘ has been launched underneath an MIT license by a developer (ClaudioDrews)….
Supertone Releases Supertonic v3: On-Device Text-to-Speech Model with 31-Language Support, Fewer Reading Failures, and Expression Tags
Supertone launched Supertonic 3, the third era of its on-device, ONNX-based text-to-speech system. Supertonic 3 ships with 31-language assist, improved studying accuracy, fewer repeat and skip failures, and v2-compatible public ONNX belongings. It is Lightning Fast, On-Device, Multilingual and Accurate TTS. What Changed from v2 to v3 Compared with Supertonic 2, Supertonic 3 reduces repeat…
