Skip to content
Computer vision

chivier/ryze-biovision-qwen3.8-27b-v6.2-fp32

specialized multimodal model for analyzing scientific figures

2 min readupdated

Ryze BioVision is built on Qwen/Qwen3.8-27B (27.78B parameters) and fine-tuned to answer four-option multiple-choice questions (A–D) based on images from biology papers. The adapter (LoRA — a lightweight add-on for a base model) is fully merged into the base weights in FP32 format, so it loads as a standalone checkpoint without extra modules. It targets tasks like color identification, comparing values along a shared axis, spatial relationships, and trend directions in plots, aiming to automate the verification of biological diagrams and charts from open-access PubMed Central sources.

Key features

  • Multimodal analysis: processes charts and figures from biology research at resolutions up to 1,048,576 pixels.
  • Four-option multiple choice: outputs strictly a single capital letter (A, B, C, or D) in response to a given question.
  • Domain specialization: trained on a dataset of PubMed Central biology papers covering colors, axes, spatial relationships, and trends.
  • FP32 format: weights are fully merged across all layers, bringing the total checkpoint size to roughly 102.7 GB.

What you can do

  • Analyze charts: test the model on trend detection, value comparisons, and matching elements in diagrams from biology publications.
  • Integrate into pipelines: run it in Python scripts via the transformers library to automate scientific figure parsing.
  • Benchmark accuracy: verify model predictions against ground-truth labels using the test splits referenced in the model card.

Example

import torch
from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "chivier/ryze-biovision-qwen3.8-27b-v6.2-fp32"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.float32,
    device_map="auto",
    max_memory={0: "60GiB", "cpu": "45GiB"},
    offload_folder="./offload",
)
model.eval()
Read moreDetails
Held-out test accuracy (383 questions): 98.17% (up from 94.26% for the base model).
Hardware requirements: the FP32 checkpoint does not fit entirely on a 96 GB GPU and requires CPU and disk offloading.
Limitations: outputs only letters A–D even when an image is unreadable; not intended for clinical or diagnostic use.
Training: 1 epoch of SFT on the Qwen3_5ForConditionalGeneration architecture with a frozen vision tower.
Discuss Telegram 0 22

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • F32weights 103 GB · ready files113 GB
  • BF16weights 52 GB57 GB
  • FP8 / INT8recommendedweights 26 GB29 GB
  • 4-bitweights 15 GB17 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # chivier/ryze-biovision-qwen3.8-27b-v6.2-fp32
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download chivier/ryze-biovision-qwen3.8-27b-v6.2-fp32 --local-dir ~/models/ryze-biovision-qwen3.8-27b-v6.2-fp32
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

chivier/ryze-biovision-qwen3.8-27b-v6.2-fp32
Author
chivier
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

220

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments