Meta: Llama 3.3 70B Instruct
meta-llama/llama-3.3-70b-instruct
Cheapest provider
$0.10 / 1M
DeepInfra
Fastest provider (p95)
521 tok/s
Groq
Intelligence (composite)
58.4
MMLU-Pro · HumanEval · math · GPQA
Per-provider performance
Latency / throughput / uptime measured across providers over the last 30 minutes of live traffic. Atlas’s router weighs these per call (with the eval-gate signal) when picking a variant for Standard and Batch tiers.
| Provider | Quant | Input $/1M | Output $/1M | Latency p50 / p95 | Throughput p50 / p95 | Uptime 30m | Success |
|---|---|---|---|---|---|---|---|
| DeepInfra | q8· fp8 | $0.1000 | $0.3200 | 409ms / 3435ms | 15 / 29 tok/s | 99.06% | 99.8% |
| Inceptron | q8· fp8 | $0.1200 | $0.3800 | 586ms / 1914ms | 26 / 44 tok/s | 99.05% | 97.4% |
| AkashML | q8· fp8 | $0.1300 | $0.4000 | 537ms / 1778ms | 31.5 / 65 tok/s | 99.42% | 99.9% |
| Nebius | q8· fp8 | $0.1300 | $0.4000 | 409ms / 3175ms | 20 / 44 tok/s | 92.95% | 97.9% |
| Novita | full· bf16 | $0.1350 | $0.4000 | 539ms / 1073ms | 34 / 46 tok/s | 99.97% | 100.0% |
| Parasail | q8· int8 | $0.2200 | $0.5000 | 698ms / 1322ms | 22 / 40 tok/s | 99.96% | 80.6% |
| Cloudflare | q8· fp8 | $0.2930 | $2.2530 | 569ms / 1952ms | 18 / 39 tok/s | 97.40% | 94.5% |
| SambaNova | full· bf16 | $0.4500 | $0.9000 | 808ms / 3107ms | 34 / 203.8 tok/s | 99.54% | 97.4% |
| Groq | undisclosed | $0.5900 | $0.7900 | 295ms / 776ms | 236 / 521 tok/s | 99.61% | 96.7% |
| Friendli | undisclosed | $0.6000 | $0.6000 | — | — | — | 0.0% |
| SambaNova | full· bf16 | $0.6000 | $1.2000 | 808ms / 3107ms | 34 / 203.8 tok/s | 98.93% | 97.4% |
| WandB | full· fp16 | $0.7100 | $0.7100 | 370ms / 1123ms | 37 / 90 tok/s | 100.00% | 100.0% |
| undisclosed | $0.7200 | $0.7200 | 300ms / 1878ms | 18 / 62.6 tok/s | 88.21% | 83.8% | |
| undisclosed | $0.7200 | $0.7200 | 300ms / 1878ms | 18 / 62.6 tok/s | 97.79% | 83.8% | |
| Together | q8· fp8 | $0.8800 | $0.8800 | 1092ms / 2553ms | 32 / 78.6 tok/s | 97.56% | 97.1% |
“—” means live telemetry hasn’t accumulated enough recent traffic for that endpoint. “undisclosed” means the provider serves the model but doesn’t expose the quantization label (typically running fp8 / int8 internally).
Intelligence breakdown
Composite score is a weighted average of public benchmarks (30% MMLU-Pro, 25% code pass@1, 25% math, 20% GPQA). Numbers come from model cards and the Artificial Analysis intelligence harness; missing components are renormalised over what’s present.
MMLU-Pro
68.9
broad reasoning
Code
33.3
pass@1 (HumanEval / LiveCodeBench)
MATH
77.0
math accuracy
GPQA Diamond
50.5
hard reasoning
Source: Meta Llama 3.x cards — Llama 3.3 70B (MMLU-Pro 68.9, GPQA-Diamond 50.5, LiveCodeBench 33.3); llama.com
How Atlas routes Meta: Llama 3.3 70B Instruct
- Realtime tier — direct passthrough at the upstream’s native precision. Best for hard latency / quality guarantees.
- Standard tier — Atlas picks the cheapest provider variant whose quantization has stayed green on your operation’s eval gates. For Meta: Llama 3.3 70B Instruct that’s currently DeepInfra at $0.10/1M.
- Batch tier — async, biggest discount. Roadmapped to use provider batch APIs (OpenAI / Anthropic) where available and queued spot capacity for open-weight workloads.