NEWMEN
← back to all models

DeepSeek: R1 Distill Llama 70B

deepseek/deepseek-r1-distill-llama-70b

DeepSeekopen-weight131K context2 providersIntelligence 69.4

Cheapest provider

$0.70 / 1M

DeepInfra

Fastest provider (p95)

50 tok/s

DeepInfra

Intelligence (composite)

69.4

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.

ProviderQuantInput $/1MOutput $/1MLatency p50 / p95Throughput p50 / p95Uptime 30mSuccess
DeepInfraq8· fp8$0.7000$0.8000131ms / 1244ms48 / 50 tok/s100.00%100.0%
Novitafull· bf16$0.8000$0.8000330ms / 1005ms47 / 49 tok/s100.0%

“—” 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

69.2

broad reasoning

Code

57.5

pass@1 (HumanEval / LiveCodeBench)

MATH

94.5

math accuracy

GPQA Diamond

53.4

hard reasoning

Source: DeepSeek-R1 distill series (MATH-500 94.5, LiveCodeBench 57.5); DeepSeek-R1 technical report

How Atlas routes DeepSeek: R1 Distill Llama 70B

  • 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 DeepSeek: R1 Distill Llama 70B that’s currently DeepInfra at $0.70/1M.
  • Batch tier — async, biggest discount. Roadmapped to use provider batch APIs (OpenAI / Anthropic) where available and queued spot capacity for open-weight workloads.