NanoJev
A 0.6B decision model that plays ViZDoom, mazes and Snake without decoding
Emerged finds new products worth a look, daily — what’s new today
S
Swift-Qwen3.8-27B
Qwen3.8-27B fine-tuned to think 58% less at near-identical accuracy
Halving reasoning tokens on a 27B open model means roughly twice the throughput on the same hardware without dropping to a weaker model.
Swift-Qwen3.8-27B is a derivative of Alibaba's Qwen3.8-27B that cuts thinking tokens by 58.3% and runs about 1.95x faster on several tasks. UkisAI identified the reasoning-marker tokens that trigger overthinking loops and fine-tuned against them, rather than capping reasoning length outright. Reported scores stay within a point of the base: GPQA-Diamond 88.3%, MMLU-Pro 85.0%, LiveCodeBench v6 81.6%, AIME 2026 94%. Weights are on Hugging Face under the Swift Open License, free for research and for companies under $1M revenue, with commercial licensing above that.
More like this
qwen38-27b-rtx3090
A tuned vLLM stack that serves Qwen3.8-27B fast on one RTX 3090
Bespoke Nimble
An open 9B model, dataset and training recipe for fast typed decisions over text
InstinctFlash
A serving runtime that speeds robot policy models up for real control loops
jevman
An arcade maze turned into a latency test for decision models
Strata
Runs a 125B mixture-of-experts model on a single gaming PC, fully offline