← Back to the full write-up on hijoseromero.com

The BF16 Showdown

Qwen 3.8-27B (the new challenger) vs Qwen 3.6-35B-A3B (the reigning champ). Full precision, thinking off, tools off, one shot each. Claude Opus 4.6 rides along as the cloud wildcard, not a contestant.

Tale of the tape

Qwen 3.8-27BQwen 3.6-35BOpus 4.6 (wildcard)
File on disk54.7 GB71.1 GBcloud
Architecture27B dense35B MoE, ~3B activefrontier
Generation speed~8.8 tok/s~27 tok/s~70 tok/s
Whole exam took~58 min~14 min~40 min (max effort)
Peak memory71 GB77 GBn/a
Can see imagesYESnoyes
Cost to run the examelectricityelectricity$7.13
The rig they ran onGMKtec EVO-X2, Ryzen AI Max+ 395, 128 GB unified memoryAnthropic's cloud
Why is full precision this slow on 128 GB?

Every single token requires reading ALL of the model's active weights from memory. The dense 27B at BF16 reads ~55 GB per token; at ~256 GB/s of memory bandwidth that caps out near 5-9 tok/s no matter how much total RAM there is. The 3.6 MoE is faster despite the bigger file because only ~3B parameters are active per token (~7 GB read). More memory makes models FIT. Bandwidth makes them FAST. Quantization attacks exactly this: the quant-ladder results page runs this same exam at every quant level, and the thinking-dial page runs it at every reasoning level.

Receipts

The raw per-run metrics Qwen 3.8 · Qwen 3.6 · Opus 4.6
The exact prompts, as plain text Blockfall · Eruption · The Ledger · Blind Artist
SCOREBOARDQwen 3.8: 0  ·  Qwen 3.6: 0

These are affiliate links. If you buy through them I may earn a small commission at no extra cost to you. As an Amazon Associate I earn from qualifying purchases.