K2 Horizon 0.9B

K2-Horizon-0.9B

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

Pricing

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

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

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

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