Skip to content
Model Pareto

K2-Horizon-32B

MBZUAI Institute of Foundation ModelsOpen weightshuggingface.co ↗
Compare with nearest rivalsCompare with…
Input
—
Output
—
Speed
~81tok/sest.
Context
524Ktokens

Open weights (Apache 2.0), no first-party API; plotted at self-host estimate. Card benchmark column is labelled K2-Horizon-32B-Stage1; the repo is the only 32B release and the card says all reported results use high reasoning effort.

Rankings

Self-hostingEstimated

How we estimate →

What it would cost to run these open weights yourself on rented GPUs. No API sells this model, so this is its price on the chart.

Hardware
1× B200 180GB
FP8 weights
Throughput
~4,595 tok/s
many requests batched
Price
$0.17–1.56 /M tok
busy → light use
On your own machine
RTX 5090 32GB (int4, ~72 tok/s single-stream)
single consumer GPU or Mac
Assumptions (5)
  • FP8 weights (34.78B params) + 40% KV-cache headroom ≈ 49 GB
  • 1× B200 180GB at $5.98–$14.24/GPU-hour on-demand (2026-09-24)
  • ~4,595 output tok/s aggregate at batch 131 (bandwidth-bound)
  • Blended 3:1 input:output; prefill ~30,567 tok/s
  • Low = 75% utilization at the low GPU price; high = 20% at the high price

Benchmark scores

~ italic, dashed = no published score yet, estimated from related benchmarks.

Coding

SWE-bench Pro (Public)Estimated
~last of 19 (estimated)below measured range (<43.8%)
  1. #1Claude Opus 5.589.9%
  2. #2Claude Sonnet 5.581.3%
  3. #18Nex-N2.5-Mini43.8%
  4. ~#19K2-Horizon-32BEstimated<43.8%
No published score — estimated from related benchmarks
DeepSWE v1.1Estimated
~#31 of 32 (estimated)~34.5%
  1. #1GPT-6.1 Sol75.2%
  2. #2DeepSeek V4.1 Flash74.2%
  3. #30Nex-N2.5-Mini36.1%
  4. ~#31K2-Horizon-32BEstimated34.5%
  5. #31Hy328.0%
No published score — estimated from related benchmarks
Terminal-Bench 4.0Estimated
~#30 of 85 (estimated)~3.9%
  1. #1Claude Sonnet 5.563.6%
  2. #2Claude Opus 5.559.6%
  3. #29Gemini 3.1 Pro4.0%
  4. ~#30K2-Horizon-32BEstimated3.9%
  5. #30GPT-5.4 mini2.0%
No published score — estimated from related benchmarks
AA Coding Agent IndexEstimated
~last of 15 (estimated)below measured range (<41.1)
  1. #1Claude Opus 5.566.0
  2. #2Claude Fable 5.162.2
  3. #14GPT-6 Luna41.1
  4. ~#15K2-Horizon-32BEstimated<41.1
No published score — estimated from related benchmarks

Hard reasoning

AIME (latest)Estimated
~#5 of 25 (estimated)~92.0%
  1. #1Inkling97.1%
  2. #2Muse Glimmer94.7%
  3. #4gpt-oss-120b93.4%
  4. ~#5K2-Horizon-32BEstimated92.0%
  5. #5Xing4.0-29B-A4B90.0%
No published score — estimated from related benchmarks
AA Intelligence IndexEstimated
~#40 of 78 (estimated)~23.1
  1. #1Claude Opus 5.557.6
  2. #2Claude Sonnet 5.556.0
  3. #39GPT-5.4 mini24.1
  4. ~#40K2-Horizon-32BEstimated23.1
  5. #40Nemotron 3 Ultra22.9
No published score — estimated from related benchmarks

Agents

τ²-bench (Telecom)Estimated
~#37 of 116 (estimated)~74.7%
  1. #1Step 3.7 Flash98.5%
  2. #2Gemini 3.1 Pro95.6%
  3. #36GPT-5.4 nano76.0%
  4. ~#37K2-Horizon-32BEstimated74.7%
  5. #37Nova 2.0 Lite72.8%
No published score — estimated from related benchmarks
OSWorld-VerifiedEstimated
~#10 of 12 (estimated)~69.3%
  1. #1Kimi K384.8%
  2. #2Qwen3.8 27B84.3%
  3. #9Nex-N2.5-Mini71.2%
  4. ~#10K2-Horizon-32BEstimated69.3%
  5. #10Muse Glimmer65.9%
No published score — estimated from related benchmarks
OSWorld 2.0Estimated
~last of 5 (estimated)below measured range (<47.5%)
  1. #1GLM-5.3 Flash59.1%
  2. #2Kimi K358.3%
  3. #4GPT-5.547.5%
  4. ~#5K2-Horizon-32BEstimated<47.5%
No published score — estimated from related benchmarks
BrowseCompEstimated
~#15 of 18 (estimated)~68.1%
  1. #1Atria Dawn Preview92.5%
  2. #2GPT-6 Astra91.5%
  3. #14Inkling77.1%
  4. ~#15K2-Horizon-32BEstimated68.1%
  5. #15Mistral Medium 3.548.6%
No published score — estimated from related benchmarks
GDPval-AA (Elo)Estimated
~#25 of 87 (estimated)~1342 Elo
  1. #1Claude Opus 5.51846
  2. #2Claude Fable 5.11735
  3. #24K2 Horizon 375B A23B1349
  4. ~#25K2-Horizon-32BEstimated1342
  5. #25GPT-5.51336
No published score — estimated from related benchmarks
AutomationBench-AAEstimated
~#34 of 80 (estimated)~22.8%
  1. #1Claude Opus 5.569.5%
  2. #2DeepSeek V4.1 Flash68.9%
  3. #33Quasar 438B23.9%
  4. ~#34K2-Horizon-32BEstimated22.8%
  5. #34MiniMax-M321.3%
No published score — estimated from related benchmarks