---

license: cc-by-nc-4.0

base_model:

• Altworld/Hemmingway-1

base_model_relation: quantized

pipeline_tag: image-text-to-text

tags:

• gguf
• llama.cpp
• qwen3.8
• hemmingway
• mtp

---

Hemmingway-1 GGUF

GGUF builds of Altworld/Hemmingway-1 (Qwen3.8-27B fine-tune for

everyday writing), made from darrellbest/Hemmingway-1-VL,

the same weights in the stock Qwen3.8 layout with the vision tower restored. All credit for the model goes to

Altworld.

| File | Precision | Size |

|---|---|---:|

| Hemmingway-1-BF16.gguf | BF16, lossless, incl. the MTP head (nextn_predict_layers = 1) | 54.66 GB |

| Hemmingway-1-Q8_0.gguf | Q8_0 | 29.05 GB |

| Hemmingway-1-Q6_K.gguf | Q6_K, imatrix | 22.43 GB |

| Hemmingway-1-Q4_K_M.gguf | Q4_K_M, imatrix | 16.81 GB |

| Hemmingway-1-mmproj-BF16.gguf | vision projector, BF16, lossless | 0.93 GB |

All files keep the MTP head. The imatrix came from ~100K tokens of UltraChat training conversations rendered with the

model's own chat template.

Quant quality

llama-perplexity --kl-divergence against the BF16 logits, on held-out chat (UltraChat test conversations,

~24K tokens). "Same top token" is how often the quant's most likely next token equals BF16's.

| Build | Same top token | Mean KLD | 99th pct KLD | PPL (BF16: 3.227) |

|---|---:|---:|---:|---:|

| Q8_0 | 98.9% | 0.0076 | 0.037 | 3.246 |

| Q6_K | 98.3% | 0.0116 | 0.085 | 3.218 |

| Q4_K_M | 95.9% | 0.0479 | 0.715 | 3.221 |

Every quant was also run in llama-server (CUDA). Each answered correctly and described a held-out photo through the

mmproj. All three produced the same get_weather tool call as the transformers bf16 reference.

Verification of the BF16 files

Every one of the 866 GGUF tensors was mapped back to its safetensors source and checked, independently of the

converter:

| | tensors | result |

|---|---:|---|

| copied | 266 | byte-identical |

| rearranged by the converter (V-head reorder, conv squeeze) | 384 | exactly the same values, only reordered |

| RMSNorm weights, stored as 1 + w (llama.cpp convention) | 168 | exact |

| A = -exp(A_log) (llama.cpp convention) | 48 | bit-identical to what transformers computes at runtime |

In the mmproj, 332 vision tensors hold the identical values. The 3D patch-embedding convolution is split by

llama.cpp into its two temporal halves, and both are exact.

Engines don't produce bit-identical logits (different kernels). As a control, transformers diverges from itself

(sdpa vs eager attention) on the same near-tied tokens. llama.cpp BF16 matched transformers' greedy output for

241/543 tokens before divergence, vs 214/543 for transformers-vs-transformers.

Runtime check (llama-server, CUDA): correct text answers, and an accurate image description through the mmproj.

Use

llama-server -m Hemmingway-1-BF16.gguf --mmproj Hemmingway-1-mmproj-BF16.gguf --jinja -ngl 99

It's a thinking model. Pass "chat_template_kwargs": {"enable_thinking": false} to answer directly.

Converted with llama.cpp bd4f514 (+ a one-line fix for mtp_num_hidden_layers: null, not needed for this model).

Licence

CC BY-NC 4.0, as Hemmingway-1: non-commercial use, with credit

to Hemmingway-1 / Altworld. Commercial use: luka@hemmingway.io.