MBZUAI's IFM Releases K2 Horizon: Six Fully Open Models from 0.9B to 375B

On September 3, MBZUAI's Institute of Foundation Models (IFM) released the K2 Horizon family: six models from 0.9B to 375B parameters, all under Apache 2.0, under the slogan "Frontier Performance, Radically Open." The flagship K2-Horizon-375B (MoE, A23B) scored 47 on the Artificial Analysis Intelligence Index—about 30 points above K2 Think V2 per @ArtificialAnlys—while @Profanion relayed official claims of parity with Gemini 3.1 Pro. With weights, training code, data recipes, and checkpoints all released, multiple posters called it the largest fully open release in AI history; independent evaluations have not yet appeared.

Confirmed

• Six models on Hugging Face, trending: 0.9B, 3.7B, 7B, 32B dense, 36B MoVA (A4B), and 375B MoE (A23B), with GGUF formats; the 36B MoVA combines MoE with "Mixture-of-Values" attention, native 512K context, 4B activated per token.
• Openness extends beyond weights: training code, data construction recipes, intermediate checkpoints, fine-grained training logs, and evaluation results are public; @drcintas described it as "360-degree open" and fully reproducible.
• Officials claim top-tier coding and agentic performance at every size, with SOTA for 0.9B/3.7B/7B in their classes, spanning smartwatch to data-center deployments.
• vLLM announced day-0 support; Cerebras is a launch partner.
• Per @HongyiWang10's reposts, core contributors include students Shicheng Wen and Chufan Shi; the team teased open-sourcing its xLLM pretraining infrastructure (not yet delivered) and further post-training iterations for the 375B model.

Unconfirmed

• SOTA and "matches Gemini 3.1 Pro" claims come from the release and reposts; no independent benchmarks appear in the posts, and @HongyiWang10 noted "unverified, treat with caution."

Why it matters

Multiple authors called this the largest fully open release in AI history: the complete pipeline from training code to data recipes is public, and a future xLLM release would push transparency further. Successful reproduction or continued training by the community would lower barriers to frontier research; vLLM day-0 support and the Cerebras partnership ease deployment.

2026-09-03 ~ 2026-09-05 · 30 related posts

• MBZUAI releases K2 Horizon: six fully open models from 0.9B to 375B with training code and data — kimmonismus · 2026-09-03
• IFM open-sources K2 Horizon models from 0.9B to 375B with training code and data recipes — testingcatalog · 2026-09-03
• [source] MBZUAI's open-weights K2 Horizon 375B jumps 30 points to 47 on the Intelligence Index — ArtificialAnlys · 2026-09-03
• [source] IFM releases K2-Horizon-MoVA-36B-A4B: 4B active params with 512K context — jacek2023 · 2026-09-03
• IFM releases K2 Horizon: six fully open models from 0.9B to 375B with 512K context — rupspace · 2026-09-03
• ifm.ai releases K2 Horizon: frontier performance, radically open — Few_Painter_5588 · 2026-09-03
• K2-Horizon open-weight family (3.7B-375B, 512K ctx, Apache-2.0) gets day-0 vLLM support — HongyiWang10 · 2026-09-03
• IFM open-sources K2 Horizon fleet: six LLMs from 0.9B to a 375B MoE — tw_killian · 2026-09-04
• IFM Debuts Six '360-Open' LLMs From 0.9B to 375B With Full Training Pipeline — dr_cintas · 2026-09-04
• IFM releases series of top-performing truly open-source models, pre-training infra xLLM to follow — HongyiWang10 · 2026-09-04
• IFM open-sources K2 Horizon: MoVA architecture, 0.9B-375B, all weights and training data — solyarisoftware · 2026-09-04
• IFM's 36B MoE model K2-Horizon-MoVA-36B-A4B trends on Hugging Face — IFM · 2026-09-04
• Cerebras Is Launch Partner for K2 Horizon: Six Open-Weight Apache 2.0 Models — HongyiWang10 · 2026-09-04
• [source] MBZUAI Releases Fully Open-Source K2 Horizon, Claims Performance Par with Gemini 3.1 Pro — Profanion · 2026-09-04
• IFM's K2-Horizon-MoVA-36B MoE open-weights model trends on Hugging Face — IFM · 2026-09-05
• Open-weights MoE K2-Horizon-375B-A23B drops: 375B total, 23B active params — Scobleizer · 2026-09-05

14 near-duplicate retellings: HongyiWang10 · HongyiWang10 · HongyiWang10 · ethanCaballero · HongyiWang10 · HongyiWang10 · HongyiWang10 · HongyiWang10 · HongyiWang10 · aliscodes · HongyiWang10 · YiMaTweets · aftahi_ai · HongyiWang10