Skip to the document
Madhuopen lab

AI Insights: Latency & Cost Optimization · part 4 of 17

insight24 July 2026cost: part of ~$10 baseline phase

Open models, raw: Qwen vs Llama vs DeepSeek

Three self-hosted models on the unmodified production prompt, vLLM on Azure A100

The question

Can a self-hosted open model match GPT-5.1's quality on the exact production prompt, with no modifications?

insight

No open model matches GPT-5.1 out of the box — but the failures look like missing information, not missing capability. Qwen is the one worth fixing.

Qwen 30B-A3B5–6 / 10fastest + cheapest: 5 s, TTFT 115 ms
Llama 3.3-70B5–6 / 103–4x Qwen's cost, no quality win
DeepSeek R1-32B6 / 1015–19 s reasoning overhead
GPT-5.1 bar8.5–8.7clear winner on raw prompt

Results (median of 3 runs, GPT-5.1 judge)

Results (median of 3 runs, GPT-5.1 judge)
ModelSingle callParallelTotal timeCost/answerFailure mode
Qwen3-30B-A3B-FP85 / 105 / 105–5.7 s$0.005–0.006under-generates (2–3 insights); stale limits
Llama 3.3-70B-FP86 / 105 / 1016.6–23 s$0.017–0.023slow, barely fits A100, no quality win
DeepSeek R1-Distill-Qwen-32B6 / 1015–19 s$0.015–0.020reasoning overhead; ineligible credits
GPT-5.1 (reference)8.78.58–21 s$0.009–0.018

The shared failure pattern

Every open model failed the same way: stale contribution numbers (old 401k limits instead of the current year's), invented ineligible credits, and too few insights per single call. The failures didn't depend on the model — which pointed at missing information, not missing capability. Bigger (Llama 70B) and reasoning-style (DeepSeek) models didn't fix it; they just cost more.

Qwen's mixture-of-experts design activates only a small slice of parameters per token, and the FP8 quantization fits comfortably on one A100 — making it by far the fastest and cheapest to serve (TTFT ~115 ms).

Qwen 30BLlama 70BDeepSeek 32BvLLMA100