Experiment ·
LESPORT-L1
Can an expert-streaming llama.cpp fork run the 35B deep model in a small RAM budget with judged-equivalent answers?
From the record
- What was tested
- expert-streaming fork for the 35B
- Result
- judged 0.625 vs llama-server 0.634, paired −0.010 (−0.042…+0.018); RSS 2.75–2.81 GiB; 14–17 tok/s on the Mac mini CPU; 10/10 greedy-identical vs the fork base
- Verdict
- parity: an equivalence (same quality, cheaper or transferable), not a gain
- Set
- v0.1 R450 contexts (public dev)
- Judge
- Claude Opus
- Pre-registered
- see arm report
Linked facts · 1
- An expert-streaming engine runs the 35B model in under 3 GiB of RAM on a desktop CPU with judged-equivalent answers.An expert-streaming engine runs the 35B model in under 3 GiB of RAM on a desktop CPU with judged-equivalent answers.judged 0.625 vs llama-server 0.634, paired −0.010 (−0.042…+0.018); RSS 2.75–2.81 GiB; 14–17 tok/s on the Mac mini CPU; 10/10 greedy-identical vs the fork baseSetv0.1Statuspublic devJudgeClaude Opus (judge.v1, both orders)n57 (R450 contexts)Seeds1Date2026-10-16CaveatDesktop only, experts read from a fast SSD and page cache; phone flash speed is unmeasured (analytic bound ≈ 4–6 tok/s with compute).Sourcenot in the public repositories yetOpen the data file