"90% of tokens local, 10% frontier": HF engineer's model allocation take
TheZachMueller · x · 2026-10-26
Hugging Face engineer Zach Mueller restates his stance: run 90% of your tokens on local/open-source models and reserve 10% for frontier ones. The take stems from Ahmad Osman's point that the community is often unipolar — praising Opus 5.5 as the frontier pick while dismissing local/oss advocacy, when the two approaches aren't mutually exclusive.
More from Infra
- Running DeepSeek Flash V4.1 IQ2_XS on Four R9700 GPUs: Real Numbers — poofph · 2026-10-26
- Kubernetes works for vLLM inference, but agent sandboxes needed Modal's custom stack at 1M+ concurrency — NielsRogge · 2026-10-26
- Anthropic reportedly offers Blackstone 10%+ yield on $9B junior debt for 1.5GW TPU deal — Beth_Kindig · 2026-10-26
- $21 of DeepSeek API wrote Molten: a Rust+Metal engine serving 300+ tok/s on M5 Macs — GabGarrett · 2026-10-26
- Musk: SpaceX closes $40B GPU debt deal, warns 200+ neoclouds face die-off — XFreeze · 2026-10-26
- Strata ported to Jetson AGX Orin 64GB: Qwen3.6-35B decoding pushed to ~41 tok/s — No-Setting8461 · 2026-10-26