Releases·Model Release
LessThink-Qwen3-4B: the same model, with far less thinking [P]
A developer post-trained Qwen3-4B to cut reasoning token usage by 44% while preserving its knowledge and answer style. The entire training pipeline ran on a single GPU, making it a low-cost efficiency fine-tune. The model is available for others to try.
26 engagement1 sourceWed, Sep 30, 2026, 07:19 AM
AI Overview
- A Reddit user post-trained Qwen3-4B to use 44% fewer reasoning tokens while preserving knowledge and answer style, on one GPU
- The LessThink-Qwen3-4B project page with details is at 5ivatej.com/lessthink/
Entities
Related
Releases·Model Release
UkisAI Swift-Qwen3.8-27B: -58.3% thinking tokens, 1.95x speed, xhigh accuracy kept
UkisAI post-trained Qwen3.8 27B into Swift-Qwen3.8-27B by using RL to penalize reasoning-marker tokens linked to overthinking, then fixing accuracy with on-policy distillation. The open-source model cuts thinking tokens 58.3% and speeds inference 1.95x while keeping xhigh accuracy, and community benchmarks independently confirm the gains. A thank-you post reports 100k+ downloads, #1 finetune and #9 model on HuggingFace Trending, with a Swift1.5 checkpoint and Swift Qwen3.8 Flash Next upcoming.
UkisAI post-trained Qwen3.8 27B into Swift-Qwen3.8-27B by using RL to penalize reasoning-marker tokens linked to overthinking, then fixing accuracy with on-policy distillation. The open-source model cuts thinking tokens 58.3% and speeds inference 1.95x while keeping xhigh accuracy, and community benchmarks independently confirm the gains. A thank-you post reports 100k+ downloads, #1 finetune and #9 model on HuggingFace Trending, with a Swift1.5 checkpoint and Swift Qwen3.8 Flash Next upcoming.
Updated Thu, Sep 24, 2026, 04:32 PM — Model surpassed 100k downloads, becoming #1 finetune and #9 model on HuggingFace Trending; Swift1.5 and Flash Next annou
3003.5k engagement4 sourcesredditMon, Sep 14, 2026, 03:57 PM
Developments·Paper
Wenyan Thinking: Classical Chinese Reasoning Cuts Trace Length in Qwen3-4B Under Decoding Budgets
A new arXiv paper introduces Wenyan Thinking, which fine-tunes Qwen3-4B-Thinking-2507 to emit intermediate reasoning in Classical Chinese while keeping answers short and direct. Rewritten wenyan supervision substantially reduces reasoning length and missing-answer rates under constrained decoding budgets, and wenyan outperforms a compact modern-Chinese control on the primary benchmark, showing linguistic register is a usable efficiency lever.
A new arXiv paper introduces Wenyan Thinking, which fine-tunes Qwen3-4B-Thinking-2507 to emit intermediate reasoning in Classical Chinese while keeping answers short and direct. Rewritten wenyan supervision substantially reduces reasoning length and missing-answer rates under constrained decoding budgets, and wenyan outperforms a compact modern-Chinese control on the primary benchmark, showing linguistic register is a usable efficiency lever.
0 engagement1 sourcearxivWed, Oct 14, 2026, 08:04 AM
Releases·Model Release
Zellige AI releases Tessera-27B-B-CJK, an Apache-2.0 Qwen3.8-27B coding agent that thinks in compressed CJK to cut tokens
Zellige AI released Tessera-27B-B-CJK, an open-weights Apache-2.0 finetune of Qwen3.8-27B trained to reason in a compressed CJK shorthand (mixed English, Chinese words, and symbols) while emitting plain English answers and tool calls. On 11 held-out coding-agent tasks it used 0.50-0.64x the base model's thinking tokens per turn with no loss in tool-call validity. It is aimed at agent loops where per-turn thinking costs compound; Zellige says to keep the base model for competition math and hard algorithmic code.
Zellige AI released Tessera-27B-B-CJK, an open-weights Apache-2.0 finetune of Qwen3.8-27B trained to reason in a compressed CJK shorthand (mixed English, Chinese words, and symbols) while emitting plain English answers and tool calls. On 11 held-out coding-agent tasks it used 0.50-0.64x the base model's thinking tokens per turn with no loss in tool-call validity. It is aimed at agent loops where per-turn thinking costs compound; Zellige says to keep the base model for competition math and hard algorithmic code.
99 engagement1 sourceredditWed, Oct 14, 2026, 05:35 PM
Social·Community
DIY pipeline distills DeepSeek v4.1 flash reasoning traces from opencode sessions into Qwen 3.8 via SFT/RL
A developer (cookiengineer) published a self-hosted SFT/RL pipeline that retrains Qwen3.8 on reasoning traces extracted from DeepSeek v4.1 flash session logs stored locally by opencode. Training takes about 12 hours on a single 48GB VRAM GPU using custom Vulkan computation shaders, and reuses existing local traces so it costs no tokens or money. The author reports the resulting model performs well on exploit development and compliance work.
A developer (cookiengineer) published a self-hosted SFT/RL pipeline that retrains Qwen3.8 on reasoning traces extracted from DeepSeek v4.1 flash session logs stored locally by opencode. Training takes about 12 hours on a single 48GB VRAM GPU using custom Vulkan computation shaders, and reuses existing local traces so it costs no tokens or money. The author reports the resulting model performs well on exploit development and compliance work.
3 engagement1 sourceredditSun, Oct 11, 2026, 08:59 AM
Social·Community
User grafts MTP onto Qwen3.8-27B for low quants, saving RAM on Vulkan
A user created a Hugging Face model that grafts Multi-Token Prediction (MTP) onto Qwen3.8-27B, addressing the lack of MTP in Unsloth's lowest quantizations. The graft reduces RAM usage compared to using an external MTP file, but the model's quality is degraded when thinking is disabled.
A user created a Hugging Face model that grafts Multi-Token Prediction (MTP) onto Qwen3.8-27B, addressing the lack of MTP in Unsloth's lowest quantizations. The graft reduces RAM usage compared to using an external MTP file, but the model's quality is degraded when thinking is disabled.
6 engagement1 sourceredditSun, Aug 23, 2026, 08:27 AM