VibeusGemma-4-31B-Humanlike-Chat brings the conversational style of Qwen... - Vibeus
Gemma-4-31B-Humanlike-Chat brings the conversational style of Qwen Humanlike 2.0 to Google’s official Gemma 4 31B. Short, lowercase replies without a special prompt, character cards, brief thinking, tools, code and formal emails — then back to chatting. The earlier Qwen GGUF releases have reached 136k downloads.
LessThanThreeAI trained it on one H200: 13,740 SFT examples, then 12 on-policy distillation steps. Two teachers handle conversation/characters and instructions/tools/code. Self-play reduced late-chat replies of just one or two words from 28% to 12%.
In the creator’s tests, with identical prompts, greedy decoding and thinking off, Google base → Humanlike scores were: When2Call 46→66, BFCL irrelevance 81→90, BFCL simple 95→96, MMLU-Pro 82.5→82.5. Against Qwen 2.0: LiveCodeBench 51→66, IFEval 83.4→90, IFBench 43.7→45.7. Some benchmarks still trail Google base; the full chart is attached.
A blind judge mistook Humanlike for the real person in 26% of 150 held-out chat moments; Google base scored 0%. These are the creator’s evaluations.
Apache-2.0, not abliterated. Available as GGUF IQ4_XS through BF16, plus BF16, FP8 and GPTQ-Int4 safetensors. The model card includes vLLM tool-calling instructions; an IQ4_XS browser demo is available.
