---
license: apache-2.0
base_model: google/gemma-4-31B-it
base_model_relation: finetune
library_name: transformers
pipeline_tag: text-generation
language:
  - en
  - ru
tags:
  - vllm
  - sglang
  - safetensors
  - gemma4
  - conversational
  - tool-calling
  - function-calling
  - roleplay
  - character
  - humanlike
  - on-policy-distillation
---

# Gemma-4-31B-Humanlike-Chat (BF16 safetensors)

**A 31B Gemma that texts like a person and still does the work.** No system prompt needed. This repo has the full BF16 weights for vLLM, SGLang and transformers.

Results, examples and how it was made are on the main card: [Gemma-4-31B-Humanlike-Chat-GGUF](https://huggingface.co/LessThanThreeAI/Gemma-4-31B-Humanlike-Chat-GGUF).

"Base" here always means Google's `google/gemma-4-31B-it`, the official release. It is not abliterated.

## vLLM

Tested with vLLM 0.31.0 on an RTX PRO 6000 Blackwell (96 GB):

```bash
vllm serve LessThanThreeAI/Gemma-4-31B-Humanlike-Chat --served-model-name humanlike-gemma \
  --max-model-len 16384 --language-model-only \
  --enable-auto-tool-choice --tool-call-parser gemma4 --reasoning-parser gemma4
```

Thinking is off by default, as in Google's Gemma 4 template. Send `chat_template_kwargs: {"enable_thinking": true}` per request to turn it on; the thinking comes back in its own field. It was trained with effort levels: add the line `Reasoning effort: low` or `Reasoning effort: high` to your system prompt, or no line for medium. Sampling: temperature 1.0, top_p 0.95, top_k 64 (Google's defaults in `generation_config.json`). For SGLang, point `--model-path` at this repo.

## VRAM

The weights take 57.8 GiB in vLLM. On the 96 GB test card at `--gpu-memory-utilization 0.88`, that left 22.7 GiB for the cache, 26,969 tokens. With less memory, take the [FP8](https://huggingface.co/LessThanThreeAI/Gemma-4-31B-Humanlike-Chat-FP8) or [GPTQ-Int4](https://huggingface.co/LessThanThreeAI/Gemma-4-31B-Humanlike-Chat-GPTQ-Int4) repo, or a GGUF from the main card.

## Checks

Same check as every file on the main card: 5 fresh test chats against the reference (the merged BF16 weights in vLLM, which match Google's base plus the Humanlike adapter at runtime: KL 0.0005), served on vLLM 0.31.0, one NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB).

| File | KL vs reference | Top-1 agreement | Greedy turns identical | Check fails (greedy / sampled) | Decode tok/s |
|---|---:|---:|---:|---:|---:|
| **BF16** | **reference** | **reference** | **reference** | **0 / 0** | **23** |

## Commissions

**Commissions open, DM codebottle on Discord (https://discord.com/users/320486798859960322).** I build custom finetunes like this one: characters, product voices, distillation into smaller models, and domain or use-case specific models.

## License

Apache-2.0, inherited from Google's Gemma 4 release.