Qwen3.8-Flash-Next Benchmarked Locally: First Model to Break 94% on 128GB Mac
tolitius · reddit · 2026-08-27
A developer tested Qwen3.8-Flash-Next on an M4 Max 128GB Mac Studio with oMLX and llama.cpp — the first model this year to break 94% on their custom cupel benchmark.
- The qwen4exp architecture isn't supported yet, so oMLX K/V caching had to be disabled; the full 4-bit quant takes 100GB, a tight fit
- On a mixed coding/general-knowledge/science benchmark, Qwen 3.8 27B led in coding but lost on general knowledge to Gemma 31B and Qwen 3.6
- Best quant: pipenetwork's MLX mixed-4/8bit build (perplexity 4.5286 vs 4.4708 bfloat16); Unsloth's GGUF UD-IQ4XS is weaker but smaller
The author plans to add more coding toolchains (pi/opencode) as models get too good to differentiate.
More from Models
- Nat Lambert: wave of AI releases rumored — thinky, muse spark, Fable 5.5, OpenAI — natolambert · 2026-10-15
- Musk amplifies dev's verdict: Grok Bot easier to use than Codex and Claude — elonmusk · 2026-10-15
- Rumor: New Anthropic model imminent, apparently behind a batch of high-severity Chrome CVEs — matthew_d_green · 2026-10-15
- StartLux-Decision debuts RL-trained calibrated decision models, 27B scores 72.8 — startlux-models · 2026-10-15
- Meta's Muse Code slammed for harsh quotas: 21% left after 5 hours — ThomasScialom · 2026-10-15
- Claude Code error mentions 'Fable 5.1' and 'Opus 5', hinting at unreleased models — adonis_singh · 2026-10-15