File size: 3,190 Bytes
ff03412 00a2d79 ff03412 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 | ---
license: cc-by-nc-4.0
base_model:
- Altworld/Hemmingway-1
base_model_relation: finetune
pipeline_tag: image-text-to-text
library_name: transformers
language:
- en
tags:
- qwen3.8
- hemmingway
- vision-language
- creative-writing
- tool-calling
---
# Hemmingway-1 VL
[Altworld/Hemmingway-1](https://huggingface.co/Altworld/Hemmingway-1) (a Qwen3.8-27B fine-tune for everyday writing),
converted from its text-only `Qwen3_5ForCausalLM` layout back into the stock
[Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) vision-language layout
(`Qwen3_5ForConditionalGeneration`), with Qwen3.8-27B's own vision tower restored. **Every language-model weight is
Hemmingway-1's, unchanged.** The layout is the one the Qwen3.8 tooling ecosystem expects (vLLM, llama.cpp mmproj,
Heretic, quantizers), and you can give it images again.
All credit for the model goes to [Altworld](https://hemmingway.io). This is a format conversion, nothing more.
## What changed, and proof that nothing else did
- Tensor names: `model.layers.*` → `model.language_model.layers.*` (and `embed_tokens`, `norm`). `lm_head.*` and the
multi-token-prediction head `mtp.*` are kept as-is.
- Added: the 333 `model.visual.*` tensors from Qwen/Qwen3.8-27B, plus its image/video processor configs.
- Config: Qwen3.8-27B's own `config.json`. Hemmingway-1's text config is identical to its `text_config`, field for field.
Verified:
| Check | Result |
|---|---|
| All 866 Hemmingway-1 tensors (incl. MTP), byte-compared | **byte-identical**, same dtypes |
| All 333 vision tensors vs Qwen3.8-27B | **byte-identical** |
| Logits, original vs this repo (transformers 5.17, bf16), on 5 prompts incl. a tool call and a tool result | **bit-equal** (max difference 0) |
| 48-token greedy continuations on the same prompts | **identical**, incl. the `<tool_call>` output |
The chat template (including tool calling) is Qwen3.8's, which Hemmingway-1 already uses unchanged.
## Use
```python
from transformers import AutoModelForImageTextToText, AutoTokenizer
tok = AutoTokenizer.from_pretrained("darrellbest/Hemmingway-1-VL")
model = AutoModelForImageTextToText.from_pretrained("darrellbest/Hemmingway-1-VL", dtype="auto", device_map="auto")
```
```bash
vllm serve darrellbest/Hemmingway-1-VL
```
Other formats, each tested: [GGUF](https://huggingface.co/darrellbest/Hemmingway-1-GGUF) (BF16 / Q8_0 / Q6_K / Q4_K_M + mmproj), [FP8](https://huggingface.co/darrellbest/Hemmingway-1-FP8) and [NVFP4](https://huggingface.co/darrellbest/Hemmingway-1-NVFP4) for vLLM. Uncensored: [darrellbest/Hemmingway-1-Heretic](https://huggingface.co/darrellbest/Hemmingway-1-Heretic).
**For fine-tuning**, either layout works. Altworld's original text-only repo is the simplest for text/tool-call LoRA.
Since the language weights are identical, an adapter trained on one applies to the other once its module names get
the `language_model.` prefix.
## Licence
Same as Hemmingway-1: [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Non-commercial use, with
credit to Hemmingway-1 / Altworld. Commercial use needs an agreement with Altworld (luka@hemmingway.io). The vision
tower comes from Qwen3.8-27B (Apache-2.0).
|