One-line: Sparse member of the K2 Horizon fleet — 36B MoE with Mixture-of-Values attention running 4B active per token; strongest agentic-terminal score in the sub-10B-active class (Terminal-Bench 2.1 58.6, beats 550B Nemotron 3 Ultra).
⚠ All scores are lab-reported (IFM model card). No independent benchmark coverage yet.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
—
HumanEval+
—
—
MBPP
—
—
LiveCodeBench
—
—
Aider polyglot
—
—
SWE-bench Verified
—
—
Terminal-Bench 2.1
58.6
lab-reported
SciCode
38.9
lab-reported
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
—
τ-Bench
—
—
τ³-Banking
26.8
lab-reported
ToolACE
—
—
GAIA
—
—
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
—
GPQA Diamond
80.8
lab-reported
MATH
—
—
AIME 2025
—
—
Humanity’s Last Exam (no tools)
25.2
lab-reported
AA-LCR (long-context)
66.3
lab-reported
Efficiency
Active params: 4B (total 36B; 100 MoE experts / 8 per token + 64 MoVA value-experts / 4 per token)
Context length: 524,288 tokens (512K)
VRAM (fp16 rough): ~72 GB total (BF16); FP8 ~36 GB; official GGUF builds released (6.0k downloads). FP8/GGUF makes it feasible on a single 48GB card.
What makes it notable
Best Terminal-Bench 2.1 of any ≤12B-active MoE in IFM’s comparison: 58.6 vs Nemotron 3 Super (38.6), Qwen3.6-35B-A3B (44.9), ai9stars G9v3-39A5B (32.6) — and it beats Nemotron 3 Ultra (550B/55B act, 53.9).
Mixture-of-Values (MoVA) attention: value-space experts on top of standard MoE — custom k2_horizon architecture, needs repo model code; vLLM/SGLang recipes provided.
For local agent loops this is the sweet spot of the fleet: 4B active per token at ~36GB quantized with 512K context; τ³-Banking 26.8 clears the 550B Nemotron Ultra (14.2) and Muse Glimmer-30B (23.5).