One-line: Flagship of IFM’s fully-open K2 Horizon fleet (Sept 3, 2026) — frontier-class agentic tool use at 23B active params, with weights, training data, code, and intermediate checkpoints all published; ⚠ large (375B total, ~750GB BF16).
⚠ All scores are lab-reported (IFM model card tables). No independent benchmark (Artificial Analysis / LMArena) published yet; The New Stack and independent reviewers note self-reported figures only. IFM’s launch blog also carries a self-audit that revised Terminal-Bench 2.1 from 70.2 → 66.9 after a reward-hacking/verification audit.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
—
HumanEval+
—
—
MBPP
—
—
LiveCodeBench
—
—
Aider polyglot
—
—
SWE-bench Verified
—
—
SWE Bench Pro (strict)
42.6
lab-reported
SWE-Atlas-QnA (strict)
48.4
lab-reported
Terminal-Bench 2.1
70.2 (66.9 audited)
lab-reported
SciCode
42.7
lab-reported
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
—
τ-Bench
—
—
τ³-Banking
34.0
lab-reported
Toolathlon Verified
65.3
lab-reported
MCPMark
67.7
lab-reported
Automation Bench Public
25.3
lab-reported
Apex-Agents (pass@1)
24.8
lab-reported
BrowseComp
72.8
lab-reported
WildClawBench
50.9
lab-reported
GDPVal-AA (Elo)
1,441
lab-reported
ToolACE
—
—
GAIA
—
—
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
—
GPQA Diamond
87.3
lab-reported
MATH
—
—
AIME 2025
—
—
Humanity’s Last Exam (no tools)
32.0
lab-reported
AA-LCR (long-context)
76.0
lab-reported
Efficiency
Active params: 23B (total 375B, 192 experts / 8 per token, 1 shared)
Context length: 524,288 tokens (512K, native from midtraining)
VRAM (fp16 rough): ⚠ large — ~750 GB (BF16); FP8 variant (~375GB) and 8×H200 vLLM/SGLang recipes provided. Not single-GPU; consumer use needs heavy quantization or API.
What makes it notable
Largest fully-open launch to date: weights + 15T-token pretraining data + training code + W&B logs + intermediate checkpoints (LLM360 lineage — IFM/Abu Dhabi, ex-MBZUAI).
Agentic tool use punches above its size: Toolathlon Verified 65.3 beats all listed open references incl. GLM 5.2 max (59.9); τ³-Banking 34.0 ≈ GLM 5.2 max (34.6); Terminal-Bench 2.1 70.2 beats Nemotron 3 Ultra (53.9) and Inkling xhigh (55.1) — audited 66.9 still ahead.
Trained via 5 expert RL branches (knowledge, IF, search, tool use, reasoning) merged before SFT; reasoning_effort=“high” required for stated scores.