One-line: Large dense member of the K2-Horizon fleet — 32B with 512K context; stage-1 checkpoint (final checkpoint to be released), claims GPQA Diamond 82.3 but trails Qwen3.8-27B on agentic coding.
⚠ All scores are lab-reported (IFM model card); baseline scores are from Artificial Analysis. The card itself flags this is a stage-1 checkpoint with stage-2 results “out soon” — treat numbers as interim.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
—
HumanEval+
—
—
MBPP
—
—
LiveCodeBench
—
—
Aider polyglot
—
—
SWE-bench Verified
—
—
Terminal-Bench 2.1
36.6
lab-reported
SciCode
30.2
lab-reported
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
—
τ³-Banking
22.5
lab-reported
ToolACE
—
—
GAIA
—
—
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
—
GPQA Diamond
82.3
lab-reported
HLE (no tools)
22.8
lab-reported
CritPt (frontier physics)
1.4
lab-reported
AA-LCR (long-context)
65.3
lab-reported
AA-Omniscience Accuracy
16.8
lab-reported
AA-Omniscience Non-Hallucination
58.3
lab-reported
Efficiency
Active params: 32B dense
Context length: 524,288 tokens (512K)
VRAM (fp16 rough): ~64 GB (BF16) — 2× H200 per the SGLang recipe; FP8/GGUF builds not yet in repo. Not a consumer-GPU model at full precision.
What makes it notable
Positioned as the “local deployment” dense flagship of the fleet, but stage-1 scores trail its own reference table: Terminal-Bench 2.1 36.6 vs Qwen3.8-27B 79.8 and Muse Glimmer-30B 51.7; τ³-Banking 22.5 vs Qwen3.8-27B 48.0.
Strong reasoning for the size: GPQA Diamond 82.3 (vs Granite 4.2-30B 64.4), HLE 22.8 (vs Muse Glimmer-30B 22.0).
Wait for the stage-2/final checkpoint before choosing it over Qwen3.8-27B for agent work — the current release is explicitly an intermediate.
Custom K2HorizonForCausalLM architecture; vLLM/SGLang first-party recipes; fully open program (data, W&B, checkpoints).