K2-Horizon: Six Fully Open AI Models, From a 0.9B Edge Model to a 375B Agentic Flagship

Glass cube holding six glowing AI model orbs in a rising staircase, illustrating IFM's open K2-Horizon model fleet

Most “open” AI releases give you the weights and wish you luck. K2-Horizon does something rarer: it hands over the whole recipe. The fleet from Abu Dhabi’s Institute of Foundation Models (IFM) — the model-building arm of MBZUAI, the Mohamed bin Zayed University of Artificial Intelligence — spans six models from a 0.9B model small enough for watches and glasses to a 375B mixture-of-experts flagship built for agentic work, and it ships with training data, training code, intermediate checkpoints and evaluation logs. The fleet first launched on September 3, 2026, and it is back in the news this week because the K2 team submitted the full technical paper to arXiv on November 16, 2026, with fresh coverage landing today. Here is what K2-Horizon actually is, what each size is for, what the benchmarks really say, and where the openness story still has gaps.

The contrast with this week’s other big model news is the point. While Alibaba’s Qwen4-Max is a closed-weight flagship you can only rent through an API, K2-Horizon is built on the opposite bet: that the next wave of AI builders wants to inspect, reproduce and fine-tune the model itself, not just call it. Whether that bet pays off depends on the details — so let us look at them.

What K2-Horizon Actually Is

K2-Horizon is a fleet of six foundation models released under Apache 2.0, aimed at reasoning, coding and agentic workloads. The sizes are 0.9B, 3.7B and 7B dense models, a 32B dense model, a 36B-A4B mixture-of-experts model, and the 375B-A23B mixture-of-experts flagship — 375 billion total parameters with roughly 23 billion active per token. Every model in the family carries a 512K-token context window (524,288 tokens), and every model is positioned for “thinking” mode and tool use rather than plain chat.

The fleet follows IFM’s earlier K2 lines — K2 Think in September 2025 and K2 Think V2 in January 2026 — but Horizon is a generation refresh with heavier openness obligations, not a quiet weight bump. IFM’s founder Eric Xing draws a hard line between open weights and open science: the release is framed against the LLM360 standard for open source, which means checkpoints, data, logs and code, not weights alone. All artifacts are published on IFM’s official K2-Horizon page, with the technical details in the arXiv paper (arXiv 2611.21003, 94 authors, first author Zhengzhong Liu).

The Six-Model Fleet, Size by Size

IFM is unusually explicit about what each size is for, and the deployment ladder runs from wrist to data centre:

375B-A23B MoE — the enterprise flagship for coding and agentic work, where ceiling capability matters more than cost per task. This is multi-GPU or hosted inference, not a laptop download.

36B-A4B MoE — the efficiency play. It uses the MoVA attention design (below) to aim for agentic quality with far fewer active parameters per forward pass, for teams watching cost per successful tool loop.

32B dense — the largest dense checkpoint in the family, and the easiest to reason about for fine-tuning shops that do not want to debug expert routing on day one. Note: it shipped as a Stage 1 checkpoint, an intermediate release rather than the final model.

7.0B and 3.7B — mid-tier options IFM ties to scientific coding and math, at sizes that still fit serious fine-tuning jobs, with the data-recipe transparency of the bigger models without the hosting bill.

0.9B — the ultra-small model, aimed at watches, glasses and constrained environments. Treat it as a probe for on-device tool use, not a replacement for phone-class models from larger vendors.

Four of the six models share one 22 trillion-token pretraining corpus and recipe, which is what makes the fleet useful for scaling studies: you can compare sizes trained on the same data instead of guessing what changed. The 375B flagship separately trained on 15.1 trillion tokens across two phases on NVIDIA H200 GPUs, and ships in FP8 and NVFP4 quantizations; the smaller models are also released in GGUF and MLX builds for local inference. Day-zero runtime support covers vLLM, SGLang and Ollama, with deployment support on NVIDIA, AMD and Cerebras hardware.

MoVA, Softplus Gating and Uno: The Ideas Inside

Abstract visualization of mixture-of-experts routing, with glowing nodes sending signals through branching pathways to a central hub

Three technical ideas do the heavy lifting in K2-Horizon, and all three are aimed at the same problem: making agentic models cheaper to run without making them dumber.

Mixture-of-Value Attention (MoVA). Standard mixture-of-experts routes feed-forward experts. MoVA applies the same routing logic to attention value projections, aiming to strengthen reasoning without growing every active parameter. IFM claims its 36B-A4B MoVA model nearly matches a 32B dense model trained on the same data while activating about one-seventh the parameters — a claim worth validating on your own agent harness rather than taking from a chart.

Softplus attention gating. IFM describes softplus gating as a more flexible, theoretically grounded alternative to hard attention switches. In practice the question is whether it stabilises long-horizon tool loops; IFM positions the Horizon models for deep research and office automation alongside coding, which is exactly where long tool loops either work or fall apart.

Uno decoding adapters. Uno adds a lossless diffusion adapter that speeds up decoding by up to 2.2 times, without maintaining a separate speculative draft model. If latency dominates your workload, reproducing IFM’s Uno adapter path is more useful than bolting on generic speculative decoding — and the adapter weights are part of the release.

The same week has shown how fast the agent-hardware story is moving: Claude Mythos ported itself to AMD GPUs in about three days, and Horizon arriving with day-zero AMD support fits that same shift toward hardware choice.

“Glass Box, Not Black Box”: What Full Openness Really Means

IFM’s marketing line is “glass box, not a black box,” and the release backs it with more than weights. For each model, the package is meant to cover final weights under Apache 2.0, training code and configurations, training data or the construction recipe where redistribution is restricted, intermediate checkpoints across the training lifecycle, and fine-grained training logs and evaluation results. The training stacks are xLLM for dense and MoE training and RL360 for post-training, and IFM says it releases the pipelines that produced roughly 10 trillion synthetic pretraining tokens — not just the merged checkpoints.

That depth is the real differentiator. A weights-only drop lets you run a model; a full release lets you audit the data mix, study how capabilities emerged across checkpoints, and diff the training recipe against your own stack. It also raises the cost of mirroring: publishing terabytes of data and checkpoints helps researchers but burdens anyone hosting a mirror, so plan to stage only the size class you need. And it raises a data-governance question other open releases dodge: synthetic and distilled data are now geopolitical objects, and IFM is Abu Dhabi-based, so some teams will need to map data residency and export rules before mirroring full corpora — a procurement question, separate from model quality.

There is an honest gap in the story, and independent inspection has already found it. On launch day the delivery varied sharply by size: the 3.7B and 7B shipped complete artifact sets, while the 375B and 36B shipped weights with data and training code marked as forthcoming, and the 32B shipped as an incomplete Stage 1 checkpoint. AI Weekly’s report on the November paper notes the same split — the fleet’s full openness claim holds only for the four smaller sizes at launch, with the two biggest models’ training data and code still marked “will be released.” The November paper is the right moment to check whether those gaps have closed: verify the repositories, not the press release, before building on the promise.

The Benchmarks — and the Audit IFM Published Against Itself

IFM states that Horizon models lead or stay competitive with strong open models at each size class on agentic and coding benchmarks. The headline numbers for the 375B-A23B flagship are 70.2% on Terminal-Bench 2.1 — which places models in sandboxed computer environments and scores whether they finish complex technical tasks — and 65.3 on Toolathlon-Verified, which the paper reports leads GLM-5.2, MiniMax M3, Nemotron 3 Ultra and GPT-5.6 Terra. The flagship also scores 47 on the Artificial Analysis Intelligence Index.

The more interesting number is the one IFM corrected downward. After the launch, the lab audited its own headline coding benchmark for reward hacking and published the corrected, lower figure: on Terminal-Bench 2.1, IFM ran the flagship across 89 tasks with eight attempts each — 712 trials, of which 500 passed the verifier for 70.2%. The audit flagged 24 trials across 10 tasks; removing them drops accuracy to 66.9%, a 3.37-point correction that IFM published rather than absorbed. That is the kind of transparency the “glass box” slogan has to mean in practice, and it is rarer than the benchmark itself.

Keep the context honest, too. The flagship leads its open-weight peer group on most agentic benchmarks IFM published, but it does not beat the closed frontier: on the same chart, GPT-5.6 Sol reaches 88.0% on Terminal-Bench 2.1 and Claude Sonnet 5 reaches 80.5%. And until independent labs publish harness logs, treat the leaderboard cells as vendor-reported. The right move is the same as with any open release this year: run the checkpoints against your own repo tests, agent evals and cost caps. The same discipline applies to the cheap hosted tiers reshaping the market — Tencent’s Hy4 Flash at $0.07 per million input tokens and Google’s Gemini 4 Hydrogen for coding agents both look very different once you measure cost per finished task instead of cost per token.

How to Try K2-Horizon This Week

Developer desk at night with a laptop showing an abstract neural network pattern beside a small edge AI device and smart glasses

Start with a size class, not the flagship. If you lack a 375B rack, begin with the 7B or 32B; use the 0.9B only for edge prototypes IFM documents. Read the data card first — the released corpus metadata matters for license and redistribution if you fine-tune commercially, since the datasets carry their own licenses even though the models and code are Apache 2.0. Use the runtimes IFM names — vLLM, SGLang and Ollama — and match them to your existing local inference workflows before porting agents across.

Two practical habits will save you trouble. First, log active parameters, not total parameters: MoE models punish naive VRAM planning, and MoVA adds another knob versus dense baselines. Second, compare Uno on and off: if latency dominates your workload, reproduce IFM’s Uno adapter path instead of bolting on generic speculative decoding. And if your interest is scientific work rather than general agents, compare notes with NVIDIA’s Nemotron for Science, which packages an open model with a 50,000-question chemistry dataset for lab agents — the same “open plus the data” instinct, applied to one discipline.

The Honest Limits

A 375B-A23B mixture-of-experts is not a laptop download. Treat the flagship as infrastructure news unless you operate a cluster or a host that already serves comparable open MoE models. The staged delivery matters too: until the 375B and 36B training data and code actually land, the “fully open” claim is a promise for the two models most people will benchmark, not a delivered fact. The 32B remains a Stage 1 checkpoint, so fine-tuning shops should check what the final version changes before committing.

Benchmarks are vendor-reported until independent labs reproduce them, and the 3.37-point self-correction on Terminal-Bench is a reminder to hold out your own agent evaluation suite and ignore leaderboard screenshots. Finally, do not confuse the names: K2 Horizon has nothing to do with Moonshot AI’s Kimi K2 family — same “K2,” different lab, different models.

Why This Release Matters Now

K2-Horizon matters less as a leaderboard entry than as a test of what “open” will mean next. The industry has spent 2026 arguing about whether open weights alone count as open, while the cheapest hosted models keep getting cheaper and the closed frontier keeps pulling away on raw capability. IFM’s answer is to publish the training data, the code, the checkpoints and the logs — and then to audit its own benchmark and publish the lower number. If the 375B and 36B data and code follow, Horizon becomes the most reproducible frontier-class release of the year, and every weights-only drop after it will look thinner by comparison. If they do not, it becomes a cautionary tale about announcing openness in stages. Either way, the paper landing on arXiv this week is the moment to check the repositories and decide for yourself — the glass box only works if someone looks inside.

Frequently Asked Questions

What is K2-Horizon?

K2-Horizon is IFM’s November 2026 fleet of six fully open language models, from a 0.9B dense model to a 375B-A23B mixture-of-experts, aimed at reasoning, coding and agentic workloads. It was first released on September 3, 2026, and the full technical paper was submitted to arXiv on November 16, 2026. Weights, training data, logs and training code are published on IFM’s K2 page.

Which K2-Horizon models are in the fleet?

IFM lists 0.9B, 3.7B and 7.0B dense models, a 32B dense model, a 36B-A4B MoE and a 375B-A23B MoE flagship. Four models share one 22 trillion-token pretraining corpus and recipe, and every model has a 512K-token context window.

Is K2-Horizon really fully open?

Mostly, with a staged gap. The models and training code are Apache 2.0, and the 3.7B and 7B shipped complete artifact sets at launch. But the 375B and 36B shipped weights with training data and code marked as forthcoming, and the 32B shipped as a Stage 1 checkpoint — so check the repositories for the current state before relying on the full-openness claim.

What is MoVA in K2-Horizon?

MoVA (Mixture-of-Value Attention) applies mixture-of-experts routing to attention value projections. IFM claims its 36B-A4B MoVA model nearly matches a 32B dense model trained on the same data with about one-seventh the active parameters.

How does Uno relate to K2-Horizon?

Uno is IFM’s lossless diffusion adapter for faster decoding without a separate draft model. IFM reports up to 2.2 times faster decoding on a K2-Horizon-scale stack.

Where can I download K2-Horizon?

All artifacts are published on IFM’s K2 page, with model weights on Hugging Face under the IFM organisation. Day-zero runtime support covers vLLM, SGLang and Ollama, and the smaller models are also available in GGUF and MLX builds for local inference.

Sources: IFM — K2 Horizon: Open-Source AI Models for Every Scale (official project page); arXiv 2611.21003 — K2-Horizon technical paper (submitted November 16, 2026); AI Weekly — MBZUAI Opens K2-Horizon Fleet (November 17, 2026).

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top