Skip to content
Model Pareto

K2-Horizon-32B

MBZUAI Institute of Foundation ModelsOpen weightshuggingface.co ↗
Input
—
Output
—
Speed
~92tok/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

~last of 24 (estimated)below measured range (<43.8%)
  1. #1Claude Fable 5.293.0%
  2. #2Claude Opus 5.589.9%
  3. #23Nex-N2.5-Mini43.8%
  4. ~#24K2-Horizon-32BEstimated<43.8%
No published score — estimated from related benchmarks
DeepSWE v1.1Estimated
~#44 of 46 (estimated)~25.7%
  1. #1Gemini 4 Argon78.8%
  2. #2DeepSeek V4.1 Flash 101777.0%
  3. #43Hy328.0%
  4. ~#44K2-Horizon-32BEstimated25.7%
  5. #44Nemotron 3.5 Super VL24.6%
No published score — estimated from related benchmarks
~#42 of 102 (estimated)~4.1%
  1. #1Kimi K3.1 Pro67.2%
  2. #2Claude Sonnet 5.563.6%
  3. #41Qwen3.8 27B5.6%
  4. ~#42K2-Horizon-32BEstimated4.1%
  5. #42Gemini 3.1 Pro4.0%
No published score — estimated from related benchmarks
LiveCodeBenchEstimated
~#17 of 28 (estimated)~55.1%
  1. #1Qwen3.8-Omni-Flash92.6%
  2. #2Qwen3.8 Flash-Next91.9%
  3. #16Claude Haiku 4.561.5%
  4. ~#17K2-Horizon-32BEstimated55.1%
  5. #17GLM-5.3 NVFP453.5%
No published score — estimated from related benchmarks
~#25 of 26 (estimated)~28.1
  1. #1Claude Opus 5.561.7
  2. #2Claude Opus 559.7
  3. #24GPT-6 Luna31.1
  4. ~#25K2-Horizon-32BEstimated28.1
  5. #25Muse Spark Open26.3
No published score — estimated from related benchmarks

Hard reasoning

Agents

~#11 of 15 (estimated)~67.9%
  1. #1Kimi K384.8%
  2. #2Qwen3.8 27B84.3%
  3. #10Nex-N2.5-Mini71.2%
  4. ~#11K2-Horizon-32BEstimated67.9%
  5. #11Muse 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
~#19 of 22 (estimated)~68.9%
  1. #1Atria Dawn Preview92.5%
  2. #2GPT-6 Astra91.5%
  3. #18Nemotron 3.5 Bolt 74B-A9B70.0%
  4. ~#19K2-Horizon-32BEstimated68.9%
  5. #19Mistral Medium 3.548.6%
No published score — estimated from related benchmarks
~#30 of 92 (estimated)~1359 Elo
  1. #1Claude Opus 5.51866
  2. #2Muse Wonder Full1857
  3. #29K2 Horizon 375B A23B1361
  4. ~#30K2-Horizon-32BEstimated1359
  5. #30GPT-5.51336
No published score — estimated from related benchmarks
~#39 of 85 (estimated)~23.0%
  1. #1Claude Fable 5.276.4%
  2. #2Muse Wonder Full71.3%
  3. #38Quasar 438B23.9%
  4. ~#39K2-Horizon-32BEstimated23.0%
  5. #39MiniMax-M321.3%
No published score — estimated from related benchmarks