K2 Horizon 32B

// Overview

K2 Horizon 32B is the largest dense model in K2 Horizon, the six-model fleet the Institute of Foundation Models released on 3 September 2026 alongside the 0.9B, 3.7B, 7B, 36B-A4B and 375B-A23B sizes. It is a 32-billion-parameter decoder-only model with a native 524,288-token context window carried from the midtraining stages onward, and IFM places it — together with the sparse 36B-A4B — in the fleet's local-deployment sweet spot: strong enough for demanding reasoning, coding and agentic work, small enough for a workstation.

One caveat is stated plainly on the model card: the checkpoint published as IFM/K2-Horizon-32B is K2-Horizon-32B-Stage1, the first stage of the final model's training. IFM says the final checkpoint and stage-2 results will follow, so every number below describes the stage-1 model rather than the finished one.

On IFM's published stage-1 table the model scores 82.3 on GPQA Diamond, 65.3 on AA-LCR long-context reasoning, 36.6 on Terminal-Bench 2.1, 30.2 on SciCode, 22.8 on Humanity's Last Exam without tools and 22.5 on tau3-Banking. IFM compares it against three dense open-weight peers of similar size — Qwen3.8-27B, Muse Glimmer-30B and IBM Granite 4.2 30B — where it leads Granite 4.2 30B on seven of the nine benchmark rows and trails Qwen3.8-27B on all but one. IFM's own framing on the launch post is that the 32B ranks among the top dense models below 40 billion parameters.

IFM pairs the dense 32B with the sparse 36B-A4B deliberately: the two were trained under the same conditions on the same data so that dense and sparse architectures can be compared directly, and IFM reports that the sparse model comes close to the dense one while activating roughly 4 billion parameters per token instead of 32 billion.

The weights are Apache-2.0, and IFM publishes the training lifecycle around them — training data or detailed construction recipes, training code, model configurations, intermediate checkpoints, fine-grained logs and evaluation results. Serving recipes are published for vLLM and SGLang, the latter validated on 2× H200, both using the `k2_horizon` reasoning and tool-call parsers with `reasoning_effort="high"`, `temperature=1.0` and `top_p=0.95`. IFM is a research institute the Mohamed bin Zayed University of Artificial Intelligence launched in May 2025.

// Benchmarks

This model's scores

1. GPQA Diamond (graduate-level science QA)82.3%
2. AA-LCR (long-context reasoning)65.3%
3. AA-Omniscience Non-Hallucination58.3%
4. Terminal-Bench 2.1 (agentic terminal use)36.6%
5. SciCode (scientific coding)30.2%
6. Humanity's Last Exam (without tools)22.8%
7. tau3-Banking (agentic tool use)22.5%
8. AA-Omniscience Accuracy (factual accuracy)16.8%
9. CritPt (frontier physics reasoning)1.4%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• 32B dense weights under Apache-2.0, sized for a single workstation or a 2-GPU server
• 512K-token native context window
• 82.3 on GPQA Diamond and 65.3 on AA-LCR long-context reasoning
• Published with training data recipes, training code, intermediate checkpoints and fine-grained logs
• A controlled dense counterpart to the sparse 36B-A4B, trained on the same data under the same conditions
• Shares vocabulary, interfaces and deployment tooling with the other five Horizon sizes

// Best for

• Reach for it when you want a dense open-weight model you can fine-tune and self-host without expert-routing complexity
• Reach for it for local-workstation reasoning, coding and agentic work
• Reach for it for long-document and repo-scale tasks that need the 512K-token window
• Reach for it for dense-versus-sparse research against the matched 36B-A4B
• Wait for the final checkpoint if you need settled numbers — the published weights are a stage-1 model

// How to access

// K2 Horizon — every version

The full lineage of the K2 Horizon line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.