UkisAI cuts thinking tokens 61.4% on GLM-5.3-Flash with under 0.5% accuracy loss
Secure_Recording_472 · reddit · 2026-10-30
UkisAI released two Swift 1.5 token-efficiency models. The approach: penalize tokens linked to pathological overthinking rather than attacking reasoning length directly, then recover accuracy via RL and on-policy distillation.
- Swift1.5 GLM-5.3-Flash: -61.4% thinking tokens on MMLU-PRO, average accuracy loss only -0.46% across benchmarks (each run x5 seeds: Terminal-Bench 2.1, GPQA-Diamond, AIME 2026, LiveCodeBench v6, etc.). Ships 1-8bit GGUF, W4A16 AWQ and NVFP4 quants; its Q3 GSQ-RCO looks SOTA at its size, coming after eval completes.
- Swift1.5 Ornith-9B: -33.5% thinking tokens with no accuracy loss, and 2x fewer overthinking loops so it's less likely to stall.
Both work with their base MTP or DFLASH2 heads. A Swift DeepSeek v4.1 Flash 1017 is being evaluated, and in-house 1B/4B foundation models are slated for Nov 2026. Training dataset is on Hugging Face.
More from Infra
- Crusoe closes round at $31B valuation, from Bitcoin mining to AI data centers — nathanbenaich · 2026-10-30
- AWS hits $184B ARR, up 39.3% year over year — RihardJarc · 2026-10-30
- AI boom: US data center & computer spending now 50% larger than home construction — HaydnBelfield · 2026-10-30
- Whatever Happened to SanDisk's High Bandwidth Flash? 3TB Per GPU Promise Met With Silence — Zombiecidialfreak · 2026-10-30
- Why local inference is inefficient: moving weights costs ~100x more than computing — cis_female · 2026-10-30
- Sam Altman says self-replicating data centers are 'within grasp', eyeing compute-energy-labor bottleneck — kimmonismus · 2026-10-30