# K2-Horizon-0.9B ![Organization](https://img.shields.io/badge/Org-IFM-555) ![Context](https://img.shields.io/badge/Context-128K%20tokens-1f6feb) ![Pricing](https://img.shields.io/badge/Price-Open%20weights-2da44e) ![License](https://img.shields.io/badge/License-Other-d29922) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) ![Released](https://img.shields.io/badge/Released-Sep%202026-555) > **Quick answer:** K2-Horizon-0.9B is IFM's compact dense model — a 0.9B-class model supporting up to 131,072 tokens of context (via YaRN RoPE scaling), trained with multi-teacher distillation for math, code, STEM, and instruction following. It scores 41.7% on AIME 2025 and 79.9% on HumanEval+, both ahead of Qwen3.5-0.8B and OpenBMB-1B in IFM's comparison table. Note: this model uses a non-Apache "Other" license — check the Hugging Face card before commercial use. ## At a Glance **Where K2-Horizon-0.9B leads** - Large gap versus Qwen3.5-0.8B on math: AIME 2025 41.7% vs 1.0%, AIME 2026 48.5% vs 0.2% - Strong HumanEval+ (79.9%) and MBPP+ (68.0%) for a sub-1B model - 128K context via YaRN in a compact model **Where it lags** - GPQA Diamond (27.3%) trails both Qwen3.5-0.8B (26.3%, close) and especially Qwen3.5-2B (54.9%, a larger reference model) - BFCL v4 function calling (28.0%) is close to but not ahead of Qwen3.5-0.8B (25.3%) and OpenBMB-1B (25.2%) — a narrow lead only - Non-Apache license — verify terms before commercial deployment **Best for:** edge/low-resource deployments needing strong math and code generation in a sub-1B footprint. ## What K2-Horizon-0.9B Is K2-Horizon-0.9B is the compact member of IFM's K2-Horizon family — a 0.9B-class dense model trained via multi-teacher distillation, with separate domain teachers for math/code, STEM, and instruction-following. Unlike its larger 512K-context siblings, it supports up to 131,072 tokens via YaRN RoPE scaling. Its benchmark profile shows an unusually large jump over the closest same-size reference model (Qwen3.5-0.8B) specifically on competition math, suggesting the distillation recipe was heavily weighted toward math/code teachers. ## Specifications | Field | Value | |---|---| | Organization | IFM | | Parameters | 0.9B (dense) | | Context window | 131,072 tokens (128K, via YaRN RoPE scaling) | | Architecture | Dense decoder-only, multi-teacher distilled | | License | Other — see Hugging Face model card for exact terms | | Release date | 2026-09 | | Modality | Text | ## Pricing | | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | Open weights | — | — | Open weights: free to download and self-host (subject to the model's license terms — not Apache 2.0, verify before commercial use). ## Context Window K2-Horizon-0.9B supports up to **131,072 tokens (128K)** via YaRN RoPE scaling — notably smaller than the 512K native context of its larger K2-Horizon siblings, reflecting its edge/compact-deployment focus. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [AIME 2025](/benchmarks/mathematical-association-of-america/aime-2025) | 41.7 | [IFM model card](https://huggingface.co/IFM/K2-Horizon-0.9B) | 2026-09 | | [AIME 2026](/benchmarks/mathematical-association-of-america/aime-2026) | 48.5 | [IFM model card](https://huggingface.co/IFM/K2-Horizon-0.9B) | 2026-09 | | [HMMT 2026](/benchmarks/hmmt/hmmt-2026) | 25.8 | [IFM model card](https://huggingface.co/IFM/K2-Horizon-0.9B) | 2026-09 | | [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) | 27.3 | [IFM model card](https://huggingface.co/IFM/K2-Horizon-0.9B) | 2026-09 | | [HumanEval+](/benchmarks/evalplus/humaneval-plus) | 79.9 | [IFM model card](https://huggingface.co/IFM/K2-Horizon-0.9B) | 2026-09 | | [MBPP+](/benchmarks/evalplus/mbpp-plus) | 68.0 | [IFM model card](https://huggingface.co/IFM/K2-Horizon-0.9B) | 2026-09 | | [LiveCodeBench v6](/benchmarks/livecodebench/livecodebench-v6) | 37.4 | [IFM model card](https://huggingface.co/IFM/K2-Horizon-0.9B) | 2026-09 | | [BFCL v4](/benchmarks/bfcl-team/bfcl-v4) | 28.0 | [IFM model card](https://huggingface.co/IFM/K2-Horizon-0.9B) | 2026-09 | *Scores above are reported by IFM and shown for context; not Benchgen measurements. Protocol and provenance details in IFM's Technical Appendix (linked from the model card).* ## Use K2-Horizon-0.9B via API ```python from openai import OpenAI client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY") response = client.chat.completions.create( model="IFM/K2-Horizon-0.9B", messages=[{"role": "user", "content": "Explain the result step by step."}], temperature=0.6, top_p=0.95, max_tokens=32768, extra_body={"chat_template_kwargs": {"reasoning_effort": "high"}}, ) print(response.choices[0].message.content) ``` ## Frequently Asked Questions

What is K2-Horizon-0.9B?

It's IFM's compact dense model in the K2-Horizon family: 0.9B parameters, 128K-token context via YaRN, trained with multi-teacher distillation.

What is K2-Horizon-0.9B's context window?

131,072 tokens (128K), achieved via YaRN RoPE scaling.

How much does K2-Horizon-0.9B cost?

Open weights — free to download; cost is hosting/inference only.

Is K2-Horizon-0.9B open source?

Weights are open/downloadable, but the license is listed as "Other" (not Apache 2.0) — check the Hugging Face model card for exact terms before commercial use.

--- *Specs and scores sourced from IFM's official Hugging Face model card; third-party benchmark scores attributed inline. Last updated 2026-09-03.*

K2 Horizon 0.9B — Training jobs

Training jobs