---

license: apache-2.0

language:

• en

base_model: Pageshift-Entertainment/pagestorm-research-preview-14b-full-book

base_model_relation: quantized

tags:

• gguf
• llama.cpp
• story-generation
• staged-generation
• full-book
• ministral3

pipeline_tag: text-generation

---

PageStorm Research Preview 14B Full Book — GGUF

GGUF quantizations of

Pageshift-Entertainment/pagestorm-research-preview-14b-full-book,

a ministral3 model trained to produce a full novel from a single prompt via a

staged generation pipeline.

Files

• pagestorm-research-preview-14b-full-book-Q8_0.gguf (~14 GB)
• pagestorm-research-preview-14b-full-book-Q4_K_M.gguf (~7.7 GB)

Requirements

• A llama.cpp build whose runtime supports the mistral3 architecture

(llm_build_mistral3 / LLM_ARCH_MISTRAL3). Older builds will fail to load it.

Notes

• The Q8_0 file was converted with convert_hf_to_gguf.py --outtype q8_0.
• The Q4_K_M file was quantized from a temporary BF16 GGUF exported from the

original BF16 Hugging Face checkpoint, not requantized from Q8_0.

• The source config.json needed original_max_position_embeddings changed

from 16384.0 to integer 16384 so the converter could write the int rope

KV field.

• The model uses a staged protocol with custom role headers

(<|start_header_id|>…<|stop_header_id|>) and <|eot_id|> as the stage stop

token — it is not a plain chat model. See the base model card and its

story_stage_generation.py for the prompt protocol.

• Native context is 262144; KV at that length is large — quantize the KV cache

(--cache-type-k q8_0 --cache-type-v q8_0) and/or cap --ctx-size to fit VRAM.

Attribution

Base model © Pageshift Entertainment, Apache-2.0. This repo only redistributes a

quantized copy of those weights.