- Who it helps
- People narrowing a local-model shortlist before testing their own hardware.
- Availability
- The public index and benchmark page were readable October 23. They mix estimates with a much smaller measured sample.
What the estimate assumes
The index uses a 4K context, batch size one, key/value attention memory and an activation allowance. Longer inputs or concurrent requests can change whether a model fits.
What was measured
The benchmark page lists sixteen runs across four cards. The inspected notes do not establish repeated trials, warm-up or a mean-versus-median policy. A short generation run and a perplexity score cannot predict every task.
What you can do
- Use the assumptions to shortlist a model, then test your actual context length and workload. Open reference ↗
Where the evidence stops
Hardware recommendations remain estimates; no fit, speed or answer quality was reproduced.
Sources read. Primary documentation read October 23, 2026 UTC. No model inference or complete product workflow was executed.
Sources & further reading
Web · official
Quantized model and GPU index ↗
Read the fit assumptions and estimated quality/speed presentation. No local model was run.
Quantized · page checked 2026-10-23
Web · analysis
Quantized measurement notes ↗
The small measurement sample and absent repetition details limit comparisons.
Quantized · page checked 2026-10-23
Coverage note: Independent demonstrations and reproduced results remain unverified; the source review does not establish a tested workflow.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
Quantized estimates which local models may fit a GPU and provides a separate set of measured runs.