--- title: K2 Horizon: IFM releases six open models from 0.9B to... description: The Institute of Foundation Models (IFM), based in Abu Dhabi, released K2 Horizon, a family of six Apache 2.0 licensed AI models ranging from 0.9B to 375B... canonical: https://daily.dev/posts/k2-horizon-ifm-releases-six-open-models-from-0-9b-to-375b-with-caveats-on-completeness-cdpsffphp twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: K2 Horizon: IFM releases six open models from 0.9B to 375B, with caveats on completeness | daily.dev og:description: The Institute of Foundation Models (IFM), based in Abu Dhabi, released K2 Horizon, a family of six Apache 2.0 licensed AI models ranging from 0.9B to 375B... og:url: https://daily.dev/posts/k2-horizon-ifm-releases-six-open-models-from-0-9b-to-375b-with-caveats-on-completeness-cdpsffphp og:image: https://api.daily.dev/og/posts/CDPSfFPhp.png og:image:alt: K2 Horizon: IFM releases six open models from 0.9B to 375B, with caveats on completeness og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # K2 Horizon: IFM releases six open models from 0.9B to 375B, with caveats on completeness **[Collections](https://daily.dev/sources/collections)** · 3 min read · 1 upvotes · 0 comments ## Summary The Institute of Foundation Models (IFM), based in Abu Dhabi, released K2 Horizon, a family of six Apache 2.0 licensed AI models ranging from 0.9B to 375B parameters, sized for use cases from local development to long-horizon agents. All models run on vLLM and SGLang, with the 3.7B and larger supporting 512K context and built-in reasoning and tool-calling. IFM published training code, eval code, data recipes, training logs, and intermediate checkpoints for some models, calling it the largest fully open-source model fleet. However, only the 3.7B and 7B models shipped with the full artifact set at launch; the 0.9B, 375B, 36B-A4B, and 32B models had missing components. Critics, including CellCog CEO Nitish Garg, noted the absence of compute disclosure and missing generator prompts, seed code, and filtering heuristics needed for full reproducibility. Developer discussion compared the release to Chinese labs like Qwen, GLM, Kimi, and DeepSeek, noting K2 Horizon goes further in some respects but its incomplete rollout undercuts the ## Content The Institute of Foundation Models (IFM), based in Abu Dhabi, has released K2 Horizon: six AI models ranging from 0.9B to 375B parameters. IFM is positioning this as the largest fully open-source model fleet, releasing not just weights but training code, configs, data recipes, intermediate checkpoints, training logs, and evaluations. ## What's in the family Each size has a stated purpose: - **0.9B** – local development and on-device use - **3.7B** – single-node serving - **7B** – cost-sensitive deployment - **32B** – everyday heavy workloads - **36B-A4B** – production serving experiments (sparse MoE) - **375B** – long-horizon agents The 0.9B, 3.7B, and 7B models claim state-of-the-art results in their size classes. The 36B-A4B is the interesting architectural bet: it uses IFM's new Mixture-of-Value Attention and activates only about 4B parameters per token, getting close to dense 32B performance at a fraction of the compute cost per inference. All six were trained on roughly 20 trillion tokens. Every model from 3.7B up is Apache 2.0 licensed and supports 512K context with reasoning and tool-calling checkpoints included. vLLM has day-0 support for the full family, and SGLang works too. IFM also released xLLM, the production training infrastructure used to build K2 Horizon, along with the full agentic post-training codebase including the RL pipeline. ## Where the openness claims get complicated The "fully open" framing is getting pushback, and some of it is fair. At launch, only the 3.7B and 7B models shipped with the complete artifact set. The 0.9B, 32B, 36B-A4B, and 375B models had missing components or incomplete checkpoints, with full releases promised later. That's a meaningful gap between the marketing and what's actually available on day one. CellCog CEO Nitish Garg pointed out that compute disclosure is absent entirely — no accelerator count, training hours, or cost figures. Full reproducibility also requires generator prompts, seed code, and filtering heuristics that weren't shared. These aren't minor omissions if the goal is genuine reproducibility. Developer discussion on Hacker News drew comparisons to Chinese labs like Qwen, GLM, Kimi, and DeepSeek, which release both instruct and base weights but keep training data and stacks closed. The comparison cuts both ways: K2 Horizon goes further on process transparency, but the incomplete launch artifacts undercut the "fully open" claim. The honest read is that this is more open than most, but not as open as advertised on day one. Whether the promised follow-up releases close that gap will matter more than the launch framing. ## Questions this post answers ### What model sizes did IFM release in the K2 Horizon family and what are they intended for? IFM released six models: 0.9B for local development, 3.7B for single-node serving, 7B for cost-sensitive deployment, 32B for everyday heavy workloads, 36B-A4B for production serving experiments, and 375B for long-horizon agents. All are Apache 2.0 licensed and run on vLLM and SGLang, with 3.7B and larger supporting 512K context and built-in reasoning and tool-calling. _Track new open model family releases like this on daily.dev before picking one to deploy._ ### Which K2 Horizon models actually shipped with full open-source artifacts at launch? Only the 3.7B and 7B models shipped with the complete artifact set (training code, eval code, data recipes, logs, checkpoints) at release. The 0.9B, 32B, 36B-A4B, and 375B models had missing components or incomplete checkpoints, with the remaining pieces promised for later. _Developers weighing 'fully open' claims can follow rollout gaps like this on daily.dev._ ### What information is missing from IFM's K2 Horizon release that would be needed for full reproducibility? Compute disclosure is absent, with no accelerator counts, training hours, or cost estimates shared. Full reproducibility would also require generator prompts, seed code, and filtering heuristics, none of which were included at launch. CellCog CEO Nitish Garg specifically flagged these gaps. _Anyone assessing open-model reproducibility claims can follow coverage like this on daily.dev._ ## Similar posts on daily.dev - [ThursdAI - Jul 16 - Inkling 975B open weights, Kimi K3 at 2.8T, a 27B model on a phone & Codex hits 9M](https://daily.dev/posts/thursdai---jul-16---inkling-975b-open-weights-kimi-k3-at-2-8t-a-27b-model-on-a-phone-codex-hits--hledq9pdh) · ThursdAI · 0 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#vllm](https://daily.dev/tags/vllm) [View this post on daily.dev](https://daily.dev/posts/k2-horizon-ifm-releases-six-open-models-from-0-9b-to-375b-with-caveats-on-completeness-cdpsffphp) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"K2 Horizon: IFM releases six open models from 0.9B to 375B, with caveats on completeness","url":"https://daily.dev/posts/k2-horizon-ifm-releases-six-open-models-from-0-9b-to-375b-with-caveats-on-completeness-cdpsffphp","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/k2-horizon-ifm-releases-six-open-models-from-0-9b-to-375b-with-caveats-on-completeness-cdpsffphp"},"datePublished":"2026-09-09T12:03:52.877Z","dateModified":"2026-10-14T02:23:24.536Z","description":"The Institute of Foundation Models (IFM), based in Abu Dhabi, released K2 Horizon, a family of six Apache 2.0 licensed AI models ranging from 0.9B to 375B...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/83a1fbe3f90992413e2a5382ac6ea36f?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/83a1fbe3f90992413e2a5382ac6ea36f?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/k2-horizon-ifm-releases-six-open-models-from-0-9b-to-375b-with-caveats-on-completeness-cdpsffphp","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,vllm","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"K2 Horizon: IFM releases six open models from 0.9B to 375B, with caveats on completeness"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/k2-horizon-ifm-releases-six-open-models-from-0-9b-to-375b-with-caveats-on-completeness-cdpsffphp#faq","mainEntity":[{"@type":"Question","name":"What model sizes did IFM release in the K2 Horizon family and what are they intended for?","acceptedAnswer":{"@type":"Answer","text":"IFM released six models: 0.9B for local development, 3.7B for single-node serving, 7B for cost-sensitive deployment, 32B for everyday heavy workloads, 36B-A4B for production serving experiments, and 375B for long-horizon agents. All are Apache 2.0 licensed and run on vLLM and SGLang, with 3.7B and larger supporting 512K context and built-in reasoning and tool-calling. Track new open model family releases like this on daily.dev before picking one to deploy."}},{"@type":"Question","name":"Which K2 Horizon models actually shipped with full open-source artifacts at launch?","acceptedAnswer":{"@type":"Answer","text":"Only the 3.7B and 7B models shipped with the complete artifact set (training code, eval code, data recipes, logs, checkpoints) at release. The 0.9B, 32B, 36B-A4B, and 375B models had missing components or incomplete checkpoints, with the remaining pieces promised for later. Developers weighing 'fully open' claims can follow rollout gaps like this on daily.dev."}},{"@type":"Question","name":"What information is missing from IFM's K2 Horizon release that would be needed for full reproducibility?","acceptedAnswer":{"@type":"Answer","text":"Compute disclosure is absent, with no accelerator counts, training hours, or cost estimates shared. Full reproducibility would also require generator prompts, seed code, and filtering heuristics, none of which were included at launch. CellCog CEO Nitish Garg specifically flagged these gaps. Anyone assessing open-model reproducibility claims can follow coverage like this on daily.dev."}}]} ```