Skip to content
Model Pareto

Ternary Bonsai 2 27B

PrismMLOpen weightsternary (PrismML PTQ1_0)quantized from Qwen3.8 27Bhuggingface.co ↗
Input
$0.075/M tok
Output
$0.50/M tok
Blended
$0.181/M tok · 3:1
Speed
~84tok/sest.
Context
262Ktokens
Cost per task
~$0.21/taskest.

Open weights (Apache-2.0). Ternary g128 quantization of Qwen3.8-27B; PTQ1_0 GGUF is 5.95 GB (PQ2_0 7.21 GB). Scores are for the quantized weights.

Rankings

Self-hostingEstimated

How we estimate →

What it would cost to run these open weights yourself on rented GPUs, compared with the API price above.

Hardware
1× B200 180GB
ternary (PrismML PTQ1_0) weights
Throughput
~6,364 tok/s
many requests batched
Price
$0.13–1.16 /M tok
busy → light use
On your own machine
RTX 5090 32GB (ternary (PrismML PTQ1_0), ~210 tok/s single-stream)
single consumer GPU or Mac
Assumptions (5)
  • TERNARY (PRISMML PTQ1_0) weights (27.36B params) + 25% KV-cache headroom ≈ 7 GB
  • 1× B200 180GB at $5.98–$14.24/GPU-hour on-demand (2026-09-24)
  • ~6,364 output tok/s aggregate at batch 182 (bandwidth-bound)
  • Blended 3:1 input:output; prefill ~38,857 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

~#3 of 21 (estimated)~84.2%
  1. #1Claude Opus 596.0%
  2. #2Claude Sonnet 585.2%
  3. ~#3Ternary Bonsai 2 27BEstimated84.2%
  4. #3Beam 1.0 501B A23B81.6%
No published score — estimated from related benchmarks
SciCodeEstimated
~#16 of 119 (estimated)~58.1%
  1. #1Claude Fable 5.267.7%
  2. #2Claude Opus 5.566.9%
  3. #15Muse Spark Open58.4%
  4. ~#16Ternary Bonsai 2 27BEstimated58.1%
  5. #16GLM-5.3 NVFP458.0%
No published score — estimated from related benchmarks
~#7 of 26 (estimated)~56.4
  1. #1Claude Opus 5.561.7
  2. #2Claude Opus 559.7
  3. #6GPT-6 Sol56.7
  4. ~#7Ternary Bonsai 2 27BEstimated56.4
  5. #7GPT-6 Astra54.7
No published score — estimated from related benchmarks

Writing & chat

Vision

MMMUEstimated
~#2 of 3 (estimated)~71.3%
  1. #1Claude Haiku 4.573.2%
  2. ~#2Ternary Bonsai 2 27BEstimated71.3%
  3. #2Llama 4 Maverick69.4%
No published score — estimated from related benchmarks
~#4 of 8 (estimated)~87.3%
  1. #1MiniMax-M391.6%
  2. #2Kimi K391.1%
  3. #2Qwen3.8 27B91.1%
  4. ~#4Ternary Bonsai 2 27BEstimated87.3%
  5. #4Gemini 3.1 Pro85.3%
No published score — estimated from related benchmarks
GDP.pdf (AA)Estimated
~#20 of 48 (estimated)~19.2%
  1. #1Muse Wonder Full32.6%
  2. #2GPT-6 Astra31.0%
  3. #19MiMo-V2.6-Pro19.2%
  4. ~#20Ternary Bonsai 2 27BEstimated19.2%
  5. #20Gemini 3.1 Pro17.8%
No published score — estimated from related benchmarks
ChartographyEstimated
~#9 of 25 (estimated)~31.9%
  1. #1Claude Fable 5.273.4%
  2. #2GPT-6 Astra71.0%
  3. #8GPT-5.6 Terra34.0%
  4. ~#9Ternary Bonsai 2 27BEstimated31.9%
  5. #9GPT-5.531.0%
No published score — estimated from related benchmarks

Hard reasoning