One-line: Medium dense member of the K2 Horizon fleet — 7B with 512K context that claims SWE-bench Verified 70.6 (lab-reported), far above every published 7–12B open-weight competitor.
⚠ All scores are lab-reported (IFM model card). The SWE-bench Verified 70.6 claim is striking and unverified by third parties — treat as a claim, not a fact, until an independent run appears.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
—
HumanEval+
—
—
MBPP
—
—
LiveCodeBench
—
—
Aider polyglot
—
—
SWE-bench Verified
70.6
lab-reported
Terminal-Bench 2.1
39.1
lab-reported
SciCode
31.6
lab-reported
HMMT Feb 2026
73.3
lab-reported
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
—
τ-Bench
—
—
τ³-Banking
25.8
lab-reported
BrowseComp
59.0
lab-reported
ToolACE
—
—
GAIA
—
—
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
—
GPQA Diamond
—
—
MATH
—
—
AIME 2025
—
—
Humanity’s Last Exam (no tools)
18.6
lab-reported
AA-LCR (long-context)
68.0
lab-reported
Efficiency
Active params: 7B dense (36 layers, GQA 32→8 KV heads)
Context length: 524,288 tokens (512K)
VRAM (fp16 rough): ~14 GB (BF16); official FP8 + GGUF builds released — runs on a single consumer GPU.
What makes it notable
SWE-bench Verified 70.6 at 7B dense would be a step-change if confirmed: comparisons on the same card are Granite 4.2-8B (47.7), Qwen3.5-9B (50.8), Gemma 4-12B (30.6). Terminal-Bench 2.1 39.1 vs Qwen3.5-9B 29.2 too.
BrowseComp 59.0 reportedly beats GPT-5 (54.9) and DeepSeek V4 Flash (53.5) on the card’s reference table — again unverified.
The diffusion adapters variant (K2-Horizon-7B-Uno, paper 2609.04010) trades a bit of quality for ~2.7× faster generation — adapter repo, not a standalone model.
Custom k2_horizon architecture; vLLM/SGLang first-party recipes. Solid local agent candidate if audited numbers hold.