AI Briefing
KOSign in

UkisAI Releases Swift GLM-5.3-Flash and Ornith 9B with Reduced Token Usage

·2026.10.30 03:54

Key point

Swift GLM-5.3-Flash achieves -61.4% thinking tokens on MMLU-PRO with -0.46% average accuracy loss across nine benchmarks.

Details

  • UkisAI released Swift GLM-5.3-Flash and Swift Ornith 9B, optimized to reduce pathological overthinking and token usage while maintaining accuracy.
  • Swift GLM-5.3-Flash shows -61.4% thinking tokens on MMLU-PRO and -0.46% average accuracy loss across Terminal-Bench 2.1, DeepSWE v1.1, GPQA-Diamond, MMLU-Pro, C-Eval, IFBench, AIME 2026, HMMT, and LiveCodeBench v6.
  • Swift Ornith 9B reduces thinking token usage by 33.5% with no accuracy loss and approximately 2x fewer overthinking loops.
  • Models are available in GGUF (1bit-8bit for GLM-5.3-Flash, 4bit-8bit for Ornith 9B), W4A16AWQ, and NVFP4 formats on Hugging Face.
  • Training involved penalizing tokens linked to overthinking and using RL and on-policy distillation to regain accuracy.
  • Upcoming releases include Swift DeepSeek v4.1 Flash 1017 and 1B/4B foundation models expected in Nov. 2026.
llmopen-sourcequantizationtoken-efficiencyglm-5.3ukisaiornith-9b
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditHemmingway-1.5: 27B Open-Weight Model Claims 90% Success Rate Against AI DetectorsRedditOpenBonsai open-sources pipeline to compress BF16 models to ~2 GB ternary GGUFsPyTorchKRLiquid AI Unveils Six 2B-Scale LFM3 ModelsPyTorchKRNanbeige4.3-33B ranks first on 17 of 20 benchmarksHacker NewsGVS5H: Zero-shot self-orchestration boosts Qwen3.8-27B to 92.4% on LiveCodeBenchLG AI ResearchLG AI Research presents LOAM at ICML 2026: LLMs design RL observation and action spaces for 3x faster learning