Overview
K2 Horizon 7B is the mid-sized 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, 32B, 36B-A4B and 375B-A23B sizes. It is a 7-billion-parameter decoder-only model with a native 524,288-token context window carried from the midtraining stages onward, and IFM aims it at phones and other on-device applications as well as ordinary single-GPU serving.
IFM's claim for the small end of the fleet is that the 0.9B, 3.7B and 7B models set a new state of the art at their respective scales. Its published table for the 7B compares it against reference models roughly its size or larger: 73.3 on HMMT Feb 2026 against 63.1 for Gemma 4-12B, 70.6 on SWE-bench Verified against 50.8 for Qwen3.5-9B and 30.6 for Gemma 4-12B, 68.0 on long-context reasoning against 65.3 for Qwen3.5-9B, and 39.1 on Terminal-Bench 2.1 against 29.2 for Qwen3.5-9B. IFM also reports 59.0 on BrowseComp, which its table places above DeepSeek V4 Flash-0423 (53.5), GPT-5 (54.9) and LongCat Flash Thinking-2601 (56.6).
IFM is candid about where the small sizes still struggle: tasks that need extensive exploration and repeated recovery, such as the harder Terminal-Bench work, remain difficult for the smallest models in the fleet. The 7B is positioned as the size where coding and multi-step workflows start to become practical, with the 3.7B and 0.9B below it aimed at lighter tool use and focused interactions.
The weights are Apache-2.0, and IFM publishes the training lifecycle around them rather than only the final checkpoint: training data or detailed construction recipes, training code, model configurations, intermediate checkpoints, fine-grained logs and evaluation results. For the 7B, IFM notes that revisions can be pinned — `main` is the default checkpoint, while `base_final` and the `mid_*_final` tags identify training stages, which is what makes the released record reproducible.
Serving recipes are published for vLLM and SGLang, both single-GPU (tensor parallelism of 1) in BF16 with the `k2_horizon` reasoning and tool-call parsers. IFM asks for `reasoning_effort="high"`, `temperature=1.0`, `top_p=0.95` and at least 32,768 output tokens so reasoning is never truncated. IFM is a research institute the Mohamed bin Zayed University of Artificial Intelligence launched in May 2025, continuing the fully-open line of work it began with the 2023 LLM360 paper.
| Released | 2026-09-03 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Parameters | 7B |
| Context | 512K |
| Architecture | Dense decoder-only transformer with a 7B core. It shares core architecture, vocabulary, training methodology, interfaces, evaluation infrastructure and deployment tooling with the rest of the six-model Horizon fleet, and IFM trained it on the same token budget as the 3.7B, 32B and 36B-A4B models. |
| Modalities | Text |
| Status | Generally available |
Benchmarks
IFM's published comparison, transcribed from the model card. The card lists three reference models per row, chosen weakest-to-strongest, so a cell is blank where that model was not one of the three; the card's own extra references are Mistral Small 4 on SciCode (28.0) and Muse Glimmer-30B on tau3-Banking (24.0). Scores are percentages.
| Benchmark | K2-Horizon-7B | Gemma 4-12B | Qwen3.5-9B | Granite 4.2-8B |
|---|---|---|---|---|
| HMMT Feb 2026 (competition mathematics) | 73.3% | 63.1% | 65.7% | 66.5% |
| SWE-bench Verified (software engineering) | 70.6% | 30.6% | 50.8% | 47.7% |
| LCR (long-context reasoning) | 68% | 61.7% | 65.3% | 43.3% |
| Terminal-Bench 2.1 (agentic terminal use) | 39.1% | 27.3% | 29.2% | 18.4% |
| SciCode (scientific coding) | 31.6% | — | 27.5% | 30.4% |
| tau3-Banking (agentic tool use) | 25.8% | — | 7% | 7.6% |
| Humanity's Last Exam (expert-level reasoning) | 18.6% | 15.7% | 14.9% | 9.7% |
This model's scores
- HMMT Feb 2026 (competition mathematics)73.3%
- SWE-bench Verified (software engineering)70.6%
- LCR (long-context reasoning)68%
- BrowseComp (web browsing)59%
- Terminal-Bench 2.1 (agentic terminal use)39.1%
- SciCode (scientific coding)31.6%
- tau3-Banking (agentic tool use)25.8%
- Humanity's Last Exam (expert-level reasoning)18.6%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- 70.6 on SWE-bench Verified and 73.3 on HMMT Feb 2026 at 7B parameters
- 59.0 on BrowseComp, which IFM's table puts above several far larger models
- 512K-token native context window
- Runs on a single GPU under the published vLLM and SGLang recipes
- Apache-2.0 weights with training data recipes, code, logs and pinnable stage checkpoints
- Shares vocabulary, interfaces and deployment tooling with the other five Horizon sizes
Best for
- Reach for it for on-device and phone-class deployments that still need tool use and reasoning
- Reach for it for single-GPU self-hosting where a 30B-plus model will not fit
- Reach for it for coding and multi-step agent workflows at the small end of the fleet
- Reach for it for long-document work that needs the 512K-token window on modest hardware
- Reach for it for research on capability emergence — stage checkpoints are published and pinnable
- Look at K2 Horizon 32B or 36B-A4B instead for demanding terminal and long-horizon agentic work
How to access
| Provider | Model ID |
|---|---|
| Self-hosted (vLLM) ↗ | IFM/K2-Horizon-7B |
| Self-hosted (SGLang) ↗ | IFM/K2-Horizon-7B |
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.
| Version | Released | Context | License |
|---|---|---|---|
| K2 Horizon 375B-A23Bcurrent | 2026-09-03 | 512K | Apache-2.0 |
| K2 Horizon MoVA 36B-A4B | 2026-09-03 | 512K | Apache-2.0 |
| K2 Horizon 32B | 2026-09-03 | 512K | Apache-2.0 |
| K2 Horizon 7B | 2026-09-03 | 512K | Apache-2.0 |
FAQ
What is K2 Horizon 7B?
It is the mid-sized dense model in K2 Horizon, the six-model fleet the Institute of Foundation Models released on 3 September 2026. It has 7 billion parameters, a native 524,288-token context window and Apache-2.0 weights, and IFM aims it at phones, other on-device applications and single-GPU serving.
How good is K2 Horizon 7B for its size?
IFM says the 0.9B, 3.7B and 7B models set a new state of the art at their respective scales. On its published table the 7B scores 70.6 on SWE-bench Verified against 50.8 for Qwen3.5-9B and 47.7 for Granite 4.2-8B, 73.3 on HMMT Feb 2026 against 66.5 for Granite 4.2-8B, and 39.1 on Terminal-Bench 2.1 against 29.2 for Qwen3.5-9B. These are IFM's own reported numbers.
What does the 59.0 BrowseComp score mean?
BrowseComp measures multi-step web browsing and research. IFM reports 59.0 for the 7B and lists three reference figures above which it places: DeepSeek V4 Flash-0423 at 53.5, GPT-5 at 54.9 and LongCat Flash Thinking-2601 at 56.6. IFM notes elsewhere in the launch post that BrowseComp harnesses differ between models, and that it used the Discard-all@95k context-length setup from the DeepSeek-V3.2 technical report.
What are K2 Horizon 7B's limits?
IFM says outright that tasks needing extensive exploration and repeated recovery — the harder Terminal-Bench work — remain difficult for the fleet's smallest models. The 7B is where coding and multi-step workflows start to become practical; for demanding long-horizon agentic work IFM points at the 32B, 36B-A4B and 375B-A23B.
How do I run K2 Horizon 7B?
The Apache-2.0 weights are on Hugging Face. IFM publishes a vLLM recipe and an SGLang recipe, both single-GPU in BF16 with the k2_horizon reasoning and tool-call parsers. IFM asks for reasoning_effort="high", temperature=1.0, top_p=0.95 and at least 32,768 output tokens so reasoning is not cut off — a truncated reasoning trace is a failed response, not a shorter one.
What else is in the K2 Horizon fleet?
Six sizes share the same architecture, vocabulary and training methodology: 0.9B for constrained devices such as watches and glasses, 3.7B and this 7B for phones and on-device apps, a dense 32B and a sparse 36B-A4B for local workstations, and a 375B-A23B for enterprise deployments.