One-line: Compact dense member of the K2-Horizon fleet — 0.9B with 128K context, multi-teacher distilled (math/code/STEM/IF experts merged via MOPD), claims HumanEval+ 79.9.
⚠ All scores are lab-reported (IFM model card). IFM’s K2-Horizon family results have been backed by independent evaluation from Artificial Analysis for the fleet (per Moor Insights), but these specific 0.9B numbers are the card’s own.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
—
HumanEval+
79.9
lab-reported
MBPP
—
—
MBPP+
68.0
lab-reported
LiveCodeBench v6
37.4
lab-reported
Aider polyglot
—
—
SWE-bench Verified
—
—
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL v4
28.0
lab-reported
τ-Bench
—
—
ToolACE
—
—
GAIA
—
—
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
—
GPQA Diamond
27.3
lab-reported
AIME 2025
41.7
lab-reported
AIME 2026
48.5
lab-reported
HMMT Feb 2026
25.8
lab-reported
Efficiency
Active params: 0.9B dense
Context length: 131,072 tokens (128K, YaRN extension from 8K)
VRAM (fp16 rough): ~2 GB — runs comfortably on a Raspberry Pi-class / NPU device
What makes it notable
SOTA-in-class claims vs its own card’s reference table: Qwen3.5-0.8B (HumanEval+ 16.5), OpenBMB-1B (65.2), Qwen3.5-2B (75.6) — 79.9 HumanEval+ / 68.0 MBPP+ would put it above the 2B reference on coding.
BFCL v4 28.0 with native tool-call parser (k2_horizon XML/JSON) — functional but modest tool-use; agent utility likely from the reasoning loop plus 128K context, not raw tool accuracy.
Fully open program: training data/recipe, W&B logs, intermediate checkpoints, vLLM PR + SGLang cookbook all public.
Custom K2HorizonForCausalLM architecture — needs repo model code (trust_remote_code) or the vLLM/SGLang first-party recipes; not stock transformers.