The Known Good Updated 29 Jul 2026
Providers· Inference host· 14 tracked models

Nebius — pricing & performance

Compare models All providers Data sources
Models served
14
of 947 tracked models
Mean output speed
66
tokens per second, averaged over
13 measured endpoints
Mean latency
0.6s
time to first token, averaged over
13 measured endpoints
Cheapest blended
$0.10
per 1M tokens at a 3:1
input:output blend
Output speed on Nebius

Median output tokens per second for the 13 of 14 models on Nebius with published throughput telemetry from OpenRouter. The full table below lists all 14, measured or not. Higher is better. We do not run these measurements ourselves.

⇩
231
222
88
48
43
36
34
33
29
28
24
23
The Known Good
14 models on Nebius
Full leaderboard →
Model Creator Input $/M Output $/M Blended Speed TTFT Context
Google: Gemma 3 27B Google $0.10 $0.30 $0.15 15 1.3s 110k
Meta: Llama 3.3 70B Instruct Meta $0.13 $0.40 $0.20 24 0.6s 131k
MoonshotAI: Kimi K3 Moonshot AI $3.00 $15.00 $6.00 34 0.8s 8k
Nous: Hermes 4 405B Nous Research $1.00 $3.00 $1.50 36 0.4s 131k
Nous: Hermes 4 70B Nous Research $0.13 $0.40 $0.20 88 0.2s 131k
NVIDIA: Nemotron 3 Nano 30B A3B NVIDIA $0.060 $0.24 $0.10 28 1.2s 262k
NVIDIA: Nemotron 3 Super NVIDIA $0.30 $0.90 $0.45 222 0.9s 262k
OpenAI: gpt-oss-120b OpenAI $0.15 $0.60 $0.26 231 0.4s 131k
Qwen: Qwen2.5 VL 72B Instruct Alibaba $0.25 $0.75 $0.38 23 0.7s 32k
Qwen: Qwen3 235B A22B Instruct 2507 Alibaba $0.20 $0.60 $0.30 43 0.6s 262k
Qwen: Qwen3 30B A3B Instruct 2507 Alibaba $0.10 $0.30 $0.15 48 0.3s 262k
Qwen: Qwen3 32B Alibaba $0.10 $0.30 $0.15 33 0.5s 41k
Qwen: Qwen3 Next 80B A3B Thinking Alibaba $0.15 $1.20 $0.41 — — 128k
Z.ai: GLM 5.1 Z.ai $1.40 $4.40 $2.15 29 0.4s 203k
The Known Good
Source. Prices, endpoint availability and throughput for Nebius are ingested from OpenRouter; model metadata is enriched from models.dev. A blank cell means the figure was not published for that endpoint — it never means zero. Blended price is (3 × input + output) ÷ 4, the same 3:1 blend used everywhere on this site; see methodology. Quality scores are not provider-specific: a model scores the same wherever it is hosted, so quality lives on the model pages.