California – November 02, 2026 -- A local-inference benchmark on an AMD Strix Halo machine found that gemma-3-27b delivers a correct answer for 729 joules at 98% accuracy, while the faster, leaner trinity-nano needs 1,004 joules at 39% accuracy.
The test covered 27 local models on 190 small tasks with checkable answers, each run under two prompts, with a power meter running. Energy was priced at California summer rates of 34 cents per kilowatt-hour off-peak and 58 cents on-peak. The rank correlation between energy per token and energy per correct answer is +0.27, close to no relationship.
Token count per answer outweighs per-token efficiency
gemma-3-27b burns roughly 13x more energy per token than trinity-nano, which generates tokens 9x faster. But gemma-3-27b answers in 181 tokens, while trinity-nano needs 3,031. Cost per correct answer spans 225x from top to bottom. phi-4-reasoning sits last because four of every five attempts exhaust the token budget and return nothing.
On the easy suite, 24 of the 27 models are dominated by a cheaper model that is no worse. Only gemma-3-4b, qwen3-4b-instruct and qwen3-30b-a3b-instruct survive. The author cautions that the top three models on the hard questions also look dominated in this table.
Four models never answered wrongly in 95 tasks
Their only failures came from the 4,096-token cap cutting generation short. granite4-h-small scores 66.3% with 32 wrong answers, while gemma-4-26b-a4b scores 51.6% with none, so the pass column ranks granite 15 points higher. The penalized models are the reasoning-heavy ones.
Seconds per answer tracks joules per answer at +0.98
On this hardware, the cheap models are also the fast ones. trinity-nano is the fastest model at 72 tokens a second yet ranks twelfth from last on time to a right answer. qwen3-30b-a3b-instruct runs at less than half that rate and finishes in eight seconds.
Seven models tied on the easy suite, but their waiting times diverge: 8 seconds for qwen3-30b-a3b-instruct, 61 for nemotron-3-super-120b and 165 for devstral-2-123b. devstral-2-123b manages 2.1 tokens a second per request, measured with four requests running at once. At 98.9% accuracy, it is not usable interactively on this box.
Active parameters set per-token cost, with large exceptions
Across the set, active parameters span 123x against a 76x range in cost per token, a rank correlation of +0.77. The author read active counts from each GGUF file rather than model cards.
- llama31-8b and granite4-h-small have 8.0B and 8.8B active parameters but sit 5.2x apart in cost per token.
- gpt-oss-120b and nemotron-3-super-120b have near-identical total size, 2.2x apart in active parameters and 3.0x apart in cost per token.
- gpt-oss-120b and devstral-2-123b, again similar in total size, are 16x apart.
A check for amortized prompt cost moved the correlation only from +0.77 to +0.78, so granite is genuinely expensive per token. The box is bandwidth-bound, so cost tracks the parameters read per token, and sparse models read fewer. Nothing measured shows sparsity makes a model smarter.