GPU · Dedicated Accelerators · by comparability tier
GPU leaderboards
Ranked by capability composite inside each tier. Tiers are derived from the exact recorded GPU model, never from a label — an entry card and a datacenter part are different experiments and share no ranking.
everything above the entry cards — consumer Ampere–Ada through Blackwell / Hopper datacenter silicon, any count; ranked by capability composite
#
Model
Hardware – Engine
Ships
Capability
Tiers
Agentic
Refusals
tok/s
Date
=1T1
Qwen/Qwen3.8-27B-FP8 verified
1×NVIDIA RTX PRO 6000 Blackwell Max-Q – vllm
ndc
0.77n=89
0.95±0.027
0.94 usability
0.93 agentic work
0.97 · 0.92 · 0.83
0.92
0
67.5
Sep 11
=1T1
google/gemma-4-31B-it-qat-w4a16-ct verifiedno-think
Googleqat-w4a16
1×NVIDIA RTX PRO 6000 Blackwell Max-Q – vllm
ndc
0.82n=42
0.92±0.026
0.93 usability
0.94 agentic work
0.94 · 0.91 · 0.93
0.96
0
47.3
Sep 10
=1T1
openai/gpt-oss-120b verified
4×NVIDIA GeForce RTX 3090 – vllm
ndc
0.73n=53
0.92±0.045
0.90 usability
0.92 agentic work
0.97 · 0.84 · 0.70
0.78
0
79.2
Sep 10
=1T1
deepseek-ai/DeepSeek-V4-Flash-0731 verifiedno-think
2×NVIDIA RTX PRO 6000 Blackwell Max-Q – vllm
ndc
0.94n=50
0.89±0.048
0.89 usability
0.94 agentic work
0.95 · 0.85 · 0.69
0.97
0
57.4
Sep 10
5T2
poolside/Laguna-S-2.1-NVFP4 verifiedno-think
1×NVIDIA RTX PRO 6000 Blackwell Max-Q – vllm
ndc
0.82n=85
0.87±0.050
0.85 usability
0.87 agentic work
0.93 · 0.83 · 0.58
0.80
0
98.6
Sep 10
=1T1
0.95±0.027
capability· tied
ndc · vllm98% done · 4 timeouts67.5 tok/s
Categories:
Sep 11
=1T1
0.92±0.026
capability· tied
ndc · vllm100% done47.3 tok/s
Categories:
Sep 10
=1T1
openai/gpt-oss-120bGPU verified
OpenAI0.92±0.045
capability· tied
ndc · vllm99% done · 1 timeout79.2 tok/s
Categories:
Sep 10
=1T1
deepseek-ai/DeepSeek-V4-Flash-0731GPU verifiedno-think
DeepSeek0.89±0.048
capability· tied
ndc · vllm100% done57.4 tok/s
Categories:
Sep 10
5T2
poolside/Laguna-S-2.1-NVFP4GPU verifiedno-think
poolside0.87±0.050
capability
ndc · vllm100% done98.6 tok/s
Categories:
Sep 10
Entry GPU (16 GB class)
filter & Pareto view →single entry datacenter card — Tesla T4 / L4 class: the 'runs on what you already have' tier; ranked by usability because item budgets bite here
#
Model
Hardware – Engine
Ships
Capability
Tiers
Agentic
Refusals
tok/s
Date
1T1
hf.co/google/gemma-4-12B-it-qat-q4_0-gguf:latest⚠ verified
Googleq4
1×NVIDIA Tesla T4 – ollama
cdc
—
0.48±0.166
0.59 capability
0.41 agentic work
0.57 · 0.37 · 0.39
0.00
0
17.0
Sep 10