Qwen3.8-27B-Humanlike-Chat 2.0 — LoRA fine-tune making Qwen chat like a person, with working tool calls

Products·launched

Qwen3.8-27B-Humanlike-Chat: fine-tuned Qwen model that chats like a real person instead of an AI assistant

A fine-tune of Qwen (27B) trained to make casual, human-to-human conversation rather than polished assistant-style replies. The creator built a dataset of ~125k obfuscated real chat messages and trained the model to drop the verbose, overly helpful tone. It targets users annoyed by the 'AI assistant' vibe in local LLMs.

A fine-tune of Qwen (27B) trained to make casual, human-to-human conversation rather than polished assistant-style replies. The creator built a dataset of ~125k obfuscated real chat messages and trained the model to drop the verbose, overly helpful tone. It targets users annoyed by the 'AI assistant' vibe in local LLMs.

Products·building

Group-chat agent that learns from user reactions to improve persona

A group-chat agent that uses LLMs to maintain a consistent persona by learning from user reactions (e.g., corrections, laughter, rejection of replies). It solves the problem of persona bots drifting into generic customer-support tone after a few messages, for developers building conversational agents.

A group-chat agent that uses LLMs to maintain a consistent persona by learning from user reactions (e.g., corrections, laughter, rejection of replies). It solves the problem of persona bots drifting into generic customer-support tone after a few messages, for developers building conversational agents.

Developments·Benchmark

Altworld's Hemmingway-1 (27B Qwen3.8 fine-tune) hits ~32 tok/s on M5 Max, excels at role-play fiction — llm-bench.io

A tester ran Altworld's Hemmingway-1, a 27B fine-tune of Qwen3.8-27B built for human-like writing, on a MacBook M5 Max at ~31.8 tok/s with thinking and MTP enabled. An LLM judge scored it 94.2 on role play and narrative, narrowly below one Qwen3.8-27B run (95.35), but side-by-side story samples read noticeably more like human fiction. The verdict: strong pick for fiction and everyday writing, while base Qwen remains the better generalist or coding assistant.

A tester ran Altworld's Hemmingway-1, a 27B fine-tune of Qwen3.8-27B built for human-like writing, on a MacBook M5 Max at ~31.8 tok/s with thinking and MTP enabled. An LLM judge scored it 94.2 on role play and narrative, narrowly below one Qwen3.8-27B run (95.35), but side-by-side story samples read noticeably more like human fiction. The verdict: strong pick for fiction and everyday writing, while base Qwen remains the better generalist or coding assistant.

Releases·Tool Release

mlx-dspark v0.10.0 speeds Qwen3.8-27B up to 3× on Apple Silicon

mlx-dspark, an MLX port of DeepSeek's DSpark speculative-decoding drafters, now supports Qwen3.8-27B via RadixArk's drafter, achieving up to ~3× speedups on Apple Silicon. Benchmarks on an M4 Pro show 2.45× mean speedup for 8-bit targets, with code runs hitting 3.18×.

mlx-dspark, an MLX port of DeepSeek's DSpark speculative-decoding drafters, now supports Qwen3.8-27B via RadixArk's drafter, achieving up to ~3× speedups on Apple Silicon. Benchmarks on an M4 Pro show 2.45× mean speedup for 8-bit targets, with code runs hitting 3.18×.

Social·Community

Users report Qwen3.8-27B Q8_0 runs impressively on Strix Halo

A user reports running Qwen3.8-27B Q8_0 locally on an ROG Flow Z13 with AMD's Strix Halo APU, achieving 9-19 tok/s and successfully generating a flight simulator via an agent. The setup uses Lemonade Server, llama.cpp ROCm, and a 64GB VRAM/64GB RAM split.

A user reports running Qwen3.8-27B Q8_0 locally on an ROG Flow Z13 with AMD's Strix Halo APU, achieving 9-19 tok/s and successfully generating a flight simulator via an agent. The setup uses Lemonade Server, llama.cpp ROCm, and a 64GB VRAM/64GB RAM split.