Occamy-1.0

Occamy-1.0

An open 35B image and text model from Accio-Lab, based on Qwen3.6-35B-A3B, in BF16, GGUF, FP8, NVFP4 and MLX formats. For teams that serve a large model.

Made by Accio-Lab on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Occamy-1.0 is an open 35B image and text model from Accio-Lab, built on Qwen3.6-35B-A3B and released under Apache 2.0. The maker presents it as an open model for co-work. It reads images and text and answers in text. The same model comes in many formats, such as BF16, GGUF, FP8, NVFP4 and MLX, for different runtimes.

Who it suits: Developers and teams who want a large open model that reads images and text, and who can pick the format that fits their hardware and runtime.

Industry | Who uses it | How they use it | Example
Software | AI engineers | Serve a 35B image and text model with vLLM or SGLang | A private assistant that answers questions about screenshots
AI hobbyists | Local model fans | Run a GGUF version in llama.cpp on home hardware | A local chat with a compressed copy
Research | Model evaluators | Compare formats from BF16 down to 3-bit | Test how size changes answers
Business | Operations teams | Use a co-work style assistant on their own servers | A helper for internal documents and images

You can skip it if you need a small model, since even the smallest GGUF listed is about 11.66 GB.

Where to use it: It runs in Transformers, SGLang or vLLM for the main checkpoint, in llama.cpp for GGUF, and in mlx-lm for the MLX versions, which are text only.

• A runtime that matches the checkpoint you pick
• A GPU server for the full BF16 checkpoint, which has 70.22 GB of weights
• A Mac with mlx-lm, for the text-only MLX versions, though the card lists Mac acceptance as pending

Use case | Who | What they get back
Image questions | Developers | Text answers about a picture, with the BF16 checkpoint
Text chat | Hobbyists | Replies from a compressed GGUF or MLX copy
Format testing | Researchers | Size and runtime choices across many checkpoints

How to set it up: Choose a checkpoint, check that your hardware has room for its weights, and download it. The download command below fetches the main BF16 repository, which is large.

• Python with pip, for the download tool
• Disk space of at least 70.22 GB for the BF16 weights, or less for a compressed format
• A runtime such as vLLM, SGLang or llama.cpp

1. Install the Hugging Face download tool.
2. Download the main repository, or pick another format from the maker's collection.
3. Start your runtime on the files, as the model card describes.

pip install -U huggingface_hub
hf download Accio-Lab/occamy-1.0

How to use it: Serve the checkpoint in your runtime and send text, plus images for the formats that support vision. The card points to an interactive checkpoint explorer for choosing a format and copying deployment commands, and the full commands are on the model page.

Tool | What it does
BF16 checkpoint | The original weights, for text and vision
GGUF files | Compressed copies for llama.cpp
FP8 and NVFP4 | Smaller formats for compatible GPU servers
MLX versions | Text-only copies for mlx-lm
Checkpoint explorer | Helps you pick a format and copy its commands

Example: Download the BF16 checkpoint and serve it with vLLM, which the card checked with nine compatibility cases. Send a screenshot and ask what it shows. You get a text answer from the 35B model on your own server.

For newcomers and experts: A big model can be stored in different formats. Some keep full precision and need much memory, and others use fewer bits to fit on smaller machines. A first small try is a GGUF file in llama.cpp.

For experts, the card tracks validation per format, from vLLM compatibility checks on BF16 and FP8 to llama.cpp samples on GGUF. The NVFP4 copy is an expert-only mixed precision build. The card says Mac Metal acceptance is still pending for each MLX copy.

Start from Qwen3.6-35B-A3B
Train Occamy-1.0 with the maker's Dressage framework
Release the BF16 checkpoint
Convert to GGUF, FP8, NVFP4 and MLX
Pick the format that fits your runtime

Good to know: This page covers the checkpoint formats. The maker also publishes a technical report, which is the place to look for quality results.

• Formats: There are 17 GGUF variants from IQ2_M to Q8_0, plus APEX GGUF profiles.
• Vision: The MLX versions take text only.
• Community: Others have already made 36 quantized versions of it.

Title: Weight size in GB by format
BF16: 70.22 GB
MLX 8-bit: 36.83 GB
FP8: 36.61 GB
NVFP4: 23.93 GB
MLX 4-bit: 19.51 GB
MLX 3-bit: 15.18 GB

Title: GGUF size range in GB
Smallest variant: 11.66 GB
Largest variant: 36.90 GB

• License: Apache 2.0.
• Cost to run: none named, the source lists no API key, account or paid service, though a 35B model needs a machine with a lot of memory.
• Data it reaches: only the text and images you send to your own runtime, plus the files you download.
• Not for: small machines, or image input on the MLX versions, which take text only.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 35.1B
• Weights: about 70.2 GB in BF16; about 17.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,807
• Likes: 240
• License: apache-2.0
• Based on: Qwen/Qwen3.6-35B-A3B
• Quantized versions: 36
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, agent, agentic, co-work, tool-use, long-context, mixture-of-experts, coding, conversational

Links

• Model: https://huggingface.co/Accio-Lab/occamy-1.0
• Files: https://huggingface.co/Accio-Lab/occamy-1.0/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Accio-Lab/occamy-1.0

Then follow the usage in its model card. Nothing in this kit runs by itself.