One-line: Small dense member of the K2-Horizon fleet — 3.7B with 512K context; claims SWE-bench Verified 68.6, far above every dense sub-5B open model on the card’s table.
⚠ All scores are lab-reported (IFM model card). Fleet-level results have independent Artificial Analysis backing (per Moor Insights), but these specific numbers are the card’s own, with baseline protocols that may differ.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
—
HumanEval+
—
—
MBPP
—
—
LiveCodeBench
—
—
Aider polyglot
—
—
SWE-bench Verified
68.6
lab-reported
Terminal-Bench 2.1
25.1
lab-reported
SciCode
25.9
lab-reported
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL v4
50.9
lab-reported
τ³-Banking
17.7
lab-reported
ToolACE
—
—
GAIA
—
—
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
—
GPQA Diamond
65.4
lab-reported
HLE
12.9
lab-reported
HMMT Feb 2026
70.5
lab-reported
Efficiency
Active params: 3.7B dense
Context length: 524,288 tokens (512K)
VRAM (fp16 rough): ~7.5 GB — fits a single consumer GPU at fp16; KV cache at 512K will require quant/long-context handling
What makes it notable
SWE-bench Verified 68.6 at 3.7B dense is a step-change claim vs the card’s reference table: Qwen3.5-4B (41.2), G9v3-3B (16.4), Granite 4.2-3B (32.2), Nemotron 3 Nano-4B (1.8).
BFCL v4 50.9 is near Qwen3.5-4B (55.7) — genuinely useful tool-calling for a 3.7B, paired with a native k2_horizon tool-call parser.
512K native context on a 3.7B is unusual; the card’s HMMT Feb 2026 70.5 (vs Qwen3.5-4B 61.6) suggests strong math for the size.
Custom K2HorizonForCausalLM architecture; vLLM/SGLang first-party recipes; fully open training data/recipe with intermediate checkpoints. One of the best local-agent value plays in the fleet if verified.