OWLCold LLM inference on the GoS1 bench: TinyLlama (1.1B) at ~0.07 mWh/token, Mistral 7B at ~0.96 mWh/token β both on the pre-S30 ladder
Caveats
- Pre-S30 panel. Both `mistral` (7B v0.3) and the version of `tinyllama` used here were the models on the bench before the S30 ladder refresh (2026-05-27). Mistral 7B was retired in that refresh in favour of `mistral-nemo:12b`. The data on disk is real and these were the live measurements at the time; a re-measurement on the new panel would produce a v2 via `supersedes`.
- TinyLlama returned π‘ confidence here (n=2-poll task window; ΞW close to the noise floor). Mistral 7B returned π’. The ratio is therefore not n=1 across both rows β interpret as approximate.
- CLAUDE.md prose at the time of import cited slightly different per-token figures (Mistral 0.943 mWh/token, TinyLlama 0.061 mWh/token). Numbers here are taken from the source result files.
- Energy per token is reward-asymmetric: a model that produces a longer answer at the same per-token cost will look cheaper per token but more expensive per task. The TinyLlama-vs-Mistral ratio in this finding is per-token, not per-answer.
- T3 is the standard task category in OWL's LLM measurement protocol. Cold = no warm-up; the model is loaded freshly before measurement.
What was measured
Two cold inferences on the same T3 task, executed via Ollama and measured at the wall through the Tapo P110:
tinyllama(1.1B parameters, GGUF Q4): single T3 run, 2026-04-24mistral(7B v0.3, GGUF Q4): T3 run, 2026-05-08 (the closer-to-CLAUDE.md-prose match)
Each model was unloaded before the measurement, baseline was taken across 10 polls Γ 1 s, then the task was issued and polled at 1 Hz until completion.
Numbers, from the stored result files
| Model | mWh per token | Tokens / s | W_base / W_task | Confidence | |---|---|---|---|---| | TinyLlama 1.1B | 0.0718 | (per file) | (per file) | π‘ | | Mistral 7B | 0.9639 | 49.7 | 61.8 W / 234.2 W | π’ |
Per-token ratio: Mistral 7B / TinyLlama β 13.4Γ β the larger model uses an order of magnitude more energy per token in this measurement.
What this measurement does not establish
- Answer quality is not measured here. CLAUDE.md notes that TinyLlama produced corpus-grounded but generic answers on the RAG faithfulness probe (see the separate
rag-faithfulness-rem-questionfinding); per-token energy alone is not a complete picture. - The bench does not measure end-to-end LLM inference (no client device, no network).
- These figures predate the S30 ladder refresh and the move to the dynamic model catalog. A current measurement on
qwen3:1.7bandmistral-nemo:12bis a follow-up that would create a v2.
Read alongside
The RAG faithfulness finding (rag-faithfulness-rem-question) on the same TinyLlama version β same retrieval, smaller model generated a hallucinated answer. Energy per token and answer correctness are independent axes.