The LLM benchmark aggregator.

K2 Horizon 32B

IFM (MBZUAI) · 32B

§ 1 · Profile

What it is good at

Bars run from 50 — the average of the models each source lists — so right of the line is above par. The middle column is the gap to whoever leads that scope.

Agents50.8−17.3—/267✱0/5

Tool use52−22—/81✱0/1

Core abilities49−18.4—/204✱0/3

Long context46.1−18.5—/0✱✱

Reasoning49.1−17.8—/178✱0/4

Science49−14.5—/122✱0/1

Expert reasoning45.4−15.5—/0✱✱

Coding48.9−20.2—/164✱0/5

Agentic coding49.1−19.6—/157✱0/4

Code generation46−24—/77✱0/2

Knowledge48.3−15.2—/138✱0/2

Factuality47.5−11.5—/0✱✱

§ 2 · Head to head

What it beats, and what beats it

The same models turn up scope after scope. Counted once: where both were placed, who finished higher. Bars run right for K2 Horizon 32B, left for the other.

NVIDIA Nemotron 3 Nano 4Bwon 8 · lost 0

GPT-4o Miniwon 8 · lost 0

GPT-4.1 Nanowon 8 · lost 0

Granite 4.2 3Bwon 9 · lost 0

GPT-4.1 Miniwon 7 · lost 1

Llama 4 Maverick Instructwon 7 · lost 1

Granite 4.2 30Bwon 10 · lost 2

G9v3 3Bwon 6 · lost 2

GPT-5.4 Nanowon 6 · lost 2

Command Awon 6 · lost 2

Claude Haiku 4.5won 6 · lost 2

DeepSeek V3won 6 · lost 2

Granite 4.2 8Bwon 6 · lost 3

Nemotron 3 Superwon 8 · lost 4

Qwen3.6 35B A3Bwon 8 · lost 4

G9v3 39A5Bwon 8 · lost 4

Qwen3.5 4Bwon 5 · lost 3

Grok 4.3won 5 · lost 3

Gemini 3.5 Flash Litewon 5 · lost 3

GPT-5.4 Miniwon 5 · lost 3

Gemma 4 12Bwon 5 · lost 3

Qwen3.5 9Bwon 5 · lost 4

Qwen3.6 Pluswon 4 · lost 4

GLM-4.6won 4 · lost 4

MiniMax M3won 6 · lost 6

Inklingwon 6 · lost 6

K2 Horizon 3.7Bwon 4 · lost 5

Nemotron 3 Ultrawon 5 · lost 7

Kimi K2.6won 3 · lost 5

Kimi K2won 3 · lost 5

Qwen3.7 Maxwon 3 · lost 5

Gemini 3.1 Prowon 3 · lost 5

K2 Horizon 7Bwon 3 · lost 6

Gemma 4 31Bwon 4 · lost 8

GLM-5.3 Flashwon 2 · lost 6

DeepSeek V3.2won 2 · lost 6

GPT-4.1won 2 · lost 6

Claude Sonnet 4.5won 2 · lost 6

Qwen3 235B A22B Instructwon 2 · lost 6

Gemini 2.5 Prowon 2 · lost 6

GPT-5 Miniwon 2 · lost 6

Claude Sonnet 4won 2 · lost 6

Gemini 3 Prowon 2 · lost 6

Muse Glimmer-30Bwon 3 · lost 9

DeepSeek V4 Flashwon 2 · lost 7

Claude Opus 4.5won 2 · lost 7

O4 Miniwon 2 · lost 7

Claude Sonnet 4.6won 1 · lost 7

Muse Spark 1.1won 1 · lost 7

Gemini 3.8 Flashwon 1 · lost 7

Grok 4.5won 1 · lost 7

Qwen3.8 Maxwon 1 · lost 7

Gemini 2.5 Flashwon 1 · lost 8

DeepSeek V4.1 Flashwon 1 · lost 8

GPT-5.6 Lunawon 1 · lost 11

GLM-5.2won 1 · lost 11

Claude Sonnet 5won 1 · lost 11

Qwen3.8 27Bwon 1 · lost 11

Gemini 3.6 Flashwon 0 · lost 8

Gemini 3.5 Flashwon 0 · lost 8

Claude Opus 4.8won 0 · lost 8

Claude Opus 4.7won 0 · lost 8

GPT-5.4won 0 · lost 8

Grok 4won 0 · lost 8

Gemini 3.7 Flashwon 0 · lost 8

Grok 4.6won 0 · lost 8

GPT-5.5won 0 · lost 8

Claude Fable 5.1won 0 · lost 8

Claude Fable 5won 0 · lost 8

GPT-5.2won 0 · lost 9

GLM-5.3won 0 · lost 9

DeepSeek V4 Prowon 0 · lost 9

O3won 0 · lost 9

Claude Opus 5won 0 · lost 9

Kimi K3won 0 · lost 9

GPT-5.6 Solwon 0 · lost 9

K2 Horizon MoVA 36B A4Bwon 0 · lost 12

K2 Horizon 375B A23Bwon 0 · lost 12

GPT-5.6 Terrawon 0 · lost 12

§ 3 · Sources

Where the numbers come from

1 publication, 9 figures. Every one links to the page it was read from.

✱✱ IFM — K2 Horizon launch ↗the model’s own publisher9 measuresread 2026-09-27

tau3-Banking 22.5%Terminal-Bench 2.1 36.6%SciCode 30.2%Humanity's Last Exam (without tools) 22.8%GPQA Diamond 82.3%CritPt 1.4%AA-LCR 65.3%AA-Omniscience Accuracy 16.8%AA-Omniscience Non-Hallucination 58.3%

Badge

[![PublicAI Index](https://publicai.io/model-index/badge?model=k2-horizon-32b)](https://publicai.io/model-index/m/k2-horizon-32b)