Case TS-F30CE7754 Sept 2026release

AI

“The Institute of Foundation Models (IFM) at the Mohamed bin Zayed University of Artificial Intelligence (@mbzuai) released K2 Horizon, a fleet of six fully open AI foundation models, ranging from 0.9 billion to 375 billion parameters." (plus supporting body text on openness definition, model sizing, and a quote attributed to Hector Liu)”

Plain restatementMBZUAI's Institute of Foundation Models has published a family of six language models spanning 0.9B to 375B parameters, described by the publisher as fully open, meaning weights, code, training data and methodology are made available.

Mostly accurateConfidence High
What this verdict means →

Distortion code this site does not recognise yet: unreleased_as_released. Not collectible until the field guide has an entry.

This one checks out on the facts. On 3 September 2026, the Institute of Foundation Models at MBZUAI did release K2 Horizon, and it really is six separate models: 0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B. The weights are downloadable now on Hugging Face under an Apache 2.0 licence, the quote from Hector Liu is genuine, and the descriptions of which model suits a watch, a phone, a server or an enterprise deployment come straight from IFM's own launch page. The one thing worth knowing is that "fully open" is not yet fully delivered. IFM's own model cards say the training data and training code "will be made public", and the 32B model is currently published as an intermediate checkpoint with the final version still to come, so nobody can reproduce the training today. IFM also says some datasets cannot be republished for licensing reasons and will be described with recipes instead of released outright. Everything in the post traces back to IFM's own announcement, so the openness claim is the lab's description of itself rather than an independently verified finding.

The drift / as claimed vs as evidenced

[drifted from the evidence:] The Institute of Foundation Models [drifted from the evidence:] (IFM) at the Mohamed bin Zayed University of Artificial Intelligence (@mbzuai) released K2 Horizon, a [drifted from the evidence:] fleet of six [drifted from the evidence:] fully open AI foundation models, [drifted from the evidence:] ranging from 0.9 billion to [drifted from the evidence:] 375 billion parameters." [drifted from the evidence:] (plus supporting body text on openness definition, model sizing, and [drifted from the evidence:] a quote attributed to Hector Liu)


[added by the neutral restatement:] MBZUAI's Institute of Foundation Models [added by the neutral restatement:] has published a [added by the neutral restatement:] family of six [added by the neutral restatement:] language models [added by the neutral restatement:] spanning 0.9B to [added by the neutral restatement:] 375B parameters, [added by the neutral restatement:] described by the publisher as fully open, meaning weights, code, training data and [added by the neutral restatement:] methodology are made available.

Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.

The trace / claim to source

Where it appeared
Submitted image
⌿ Omitted qualifier
A load-bearing condition from the source quietly disappears from the claim.
unreleased_as_released
$ Marketing as evidence
Promotional material dressed up as independent proof.
Tertiary sourcetrade press
BigDATAwire / HPCwire syndication of the press release
Tertiary sourcewire distribution
Zawya syndication of the MBZUAI press release
Secondary sourcenamed-outlet journalism
The National, "UAE's AI university introduces world's largest 'fully open' models", Sept 3 2026
Secondary sourceindependent evaluator
Artificial Analysis model page for K2 Horizon 375B A23B
Secondary sourcecommentary
Datastudios analysis flagging the openness timing nuance
Secondary sourcenamed analyst commentary
Moor Insights & Strategy analyst note
Primary sourcevendor
IFM official launch blog, "Introducing K2 Horizon: Frontier Performance, Radically Open"
Primary sourcevendor
IFM/K2-Horizon-375B-A23B model card, Hugging Face
Primary sourcevendor
IFM/K2-Horizon-32B model card, Hugging Face
Primary sourcevendor
IFM/K2-Horizon-MoVA-36B-A4B model card, Hugging Face
Primary sourcevendor
IFM product page for K2
Primary sourcevendor
PR Newswire release, Sept 3 2026
● Primary source found
What is true
  • The Institute of Foundation Models is a real lab, launched by MBZUAI in May 2025, with sites in Abu Dhabi, Silicon Valley and Paris.
  • K2 Horizon was released on 3 September 2026.
  • It is six models, not a marketing grouping: 0.9B, 3.7B, 7B, 32B, 36B-A4B, 375B-A23B. The count is exact.
  • The stated range, 0.9 billion to 375 billion parameters, is exact.
  • Weights are downloadable now on Hugging Face under Apache 2.0, with vLLM and SGLang serving recipes and a hosted API through named partners.
  • The per-size positioning in the caption (watch and glasses, phones, local and on-prem, enterprise) reproduces IFM's own descriptions accurately.
  • The Hector Liu quote is authentic, correctly attributed to the director of IFM's Silicon Valley lab, and the sentence quoted is verbatim.
What is misleading
  • Omitted qualifier: the caption's gloss states that fully open models let anyone "access, inspect or adapt weights, code, training data and methodology", presented as a completed present-tense fact. The model cards for the flagship, the 32B and the 36B-A4B say training data/recipe and training code "will be made public" and that intermediate checkpoints "will be released". At the moment of the announcement, the data and training code portions of "fully open" are a commitment, not a delivered artifact. A reader is led to believe reproduction is possible today.
  • Omitted qualifier: "fully open" also does not mean every dataset is published. IFM's own wording is "training data or detailed data- construction recipes", with source descriptions and mixture recipes substituted where redistribution licenses prevent publication. Models and code are Apache 2.0, but the datasets are not under one uniform license.
  • Unreleased as released: the six-model fleet is presented as uniformly shipped. One member, the dense 32B, is currently published as a Stage 1 checkpoint with the final checkpoint and stage 2 results explicitly still to come.
  • Marketing as evidence: the caption is a near-verbatim restatement of IFM's press release, including the "fully open" designation, which is the vendor's own label for its own release. Every substantive descriptor in the post traces to one interested party. This is not an error by the poster, but it means the post carries no independent verification of the openness claim it is transmitting.
What is uncertain
  • Whether the training data and training code will be published, and on what timeline. The cards give no date.
  • Whether the published data, once released, will be complete enough to actually reproduce training, given the license-restricted portions substituted with recipes. This cannot be assessed until the artifacts land.
  • The performance and "state of the art at their respective scales" claims in IFM's materials are vendor-run and were not adjudicated here, since the caption does not assert them. Artificial Analysis has begun independent indexing but a full independent replication of IFM's benchmark table was not located.
  • The superlative used by some outlets, "world's largest fully open model in history", was not tested. It depends entirely on how "fully open" is defined and would require a comparison set that no one has published.
  • All page content was obtained through search-result extraction rather than direct page loads, so the model cards may have been updated in the hours since retrieval.
Evidence summary

The release is real and the numbers in the caption are exact. IFM's own launch blog states that it is releasing K2 Horizon, a connected fleet of six models: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B. The corresponding Hugging Face repositories exist under the IFM organization, including K2-Horizon-375B-A23B, described as the flagship of the family, a sparse Mixture-of-Experts model that stores 375B parameters and runs 23B per token, with a 512K context window, plus 32B, MoVA-36B-A4B, 7B, 3.7B and 0.9B repos and GGUF conversions. The licensing and distribution details check out. Models and code are released under Apache 2.0, and the API is live through inference partners including Compass, Cerebras, AWS, and Nebius. IFM states that the models and code are under Apache 2.0, while datasets retain their applicable licenses, including licenses such as ODC-BY. The size-to-device positioning in the caption is taken almost verbatim from IFM. The 0.9B model targets highly constrained environments like watches and glasses; 3.7B and 7B bring advanced capabilities to phones and on-device apps; a dense 32B and sparse 36B-A4B cover local hosting and on-premise servers; and the 375B-A23B tops the fleet for demanding enterprise deployments. The quote is authentic and the caption reproduces the operative sentence accurately. Per the press release, Hector Liu, director of IFM's Silicon Valley lab, said the institute is releasing an entire fleet at once, six models, from one small enough to run on a watch to a flagship built for enterprise reasoning, and that developers can start on the smallest model and scale to the flagship. The one material gap is timing on "fully open." The blog frames the lifecycle release in present tense: IFM says it is releasing intermediate checkpoints, training data or detailed data-construction recipes, open architecture, mixture compositions, training code, configurations, fine-grained logs, evaluation results, and final weights. The model cards themselves are in future tense: the flagship card says intermediate checkpoints will be released, and that training data/recipe and the training code will be made public. Independent commentary flagged this directly, noting that IFM's launch materials describe the fleet as fully open and present the training lifecycle as part of the release, while the current Hugging Face card for the 375B flagship says the final checkpoint is available and that intermediate checkpoints, data, and training code will be released. Separately, one of the six is not shipped in final form: the 32B card describes K2-Horizon-32B-Stage1, notes that the final checkpoint is to be released, and states that results are for stage 1 of the final model training with stage 2 results out soon.

Complete reasoning
As of 2026-09-04, the official-channel check resolves the claim directly: IFM's own blog, product page, press release and six Hugging Face repositories all confirm a 3 September 2026 release of exactly six models spanning 0.9B to 375B, under Apache 2.0, with weights downloadable and an API live. The count, the range, the institution, the attribution and the per-size positioning are all exact, which is why "Unverified", "Inconclusive" and "False" were rejected. "Accurate" was rejected because the caption's own definition of "fully open" describes access to training data and code as a present capability, while the model cards state those artifacts will be made public, and because one of the six is currently a Stage 1 checkpoint. "Source exists but framing is misleading" was considered and rejected because that framing gap is inherited from the vendor's own launch language rather than introduced by the post, and it does not change the operative proposition, which is that six models across that parameter range were released and are downloadable today. Confidence is High because the primary artifacts were located and read for the exact models named, with no version ambiguity.
Use this case

The reply is formatted for pasting into the thread where the claim is circulating.

Compact share page: ai.trueseeker.com/s/f30ce7754cae/mrl46Cf5Ae6pRzV3gh0j2BqXbPf

Ask this case

Answers come only from the case file above; nothing is added.

Is K2 Horizon really six separate models, and is the size range accurate?

Yes. It is six distinct models ranging from 0.9 billion to 375 billion parameters, and both the count and the range match IFM's own release exactly.

Can I actually download and use the training data and training code right now?

No. The weights are downloadable today under Apache 2.0, but IFM's own model cards say the training data and training code will be made public, meaning that part of the release is a future commitment, not something available now.

Are all six models fully finished and released in final form?

No. Five of the six appear final, but the 32B model is currently published as a Stage 1 checkpoint, with the final checkpoint and stage 2 results still to come.

Does 'fully open' mean every dataset used to train these models is published?

Not entirely. IFM says some datasets cannot be redistributed due to licensing, so for those it will publish descriptions or recipes instead of the actual data.

Has anyone independently verified IFM's claim that this is the most open model of its kind?

No. The case file notes that superlative claims about being the largest fully open model were not tested, since that would require a comparison set that has not been published, and the openness claim itself traces back only to IFM's own materials.

Similar cases on record