NVFP4 variant of Qwen 3.6 35B for reasoning and tool calling.
Open-weight model · Image and text to text
PhysicalEye-Decide-35B
by Physical AI masterset-ai/PhysicalEye-Decide-35B
PhysicalEye-Decide-35B is an open-weight model for image and text to text from Physical AI, released under Apache License 2.0. It has 34.6B parameters and a 262,144-token context. At 16-bit it needs about 83 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.
One forward pass. One calibrated decision. PhysicalEye-Decide-35B answers a question about an image by choosing among the options you give it, and returns a probability for every option. It does not write text.
Parameters34.6B
Context262,144
Weights69.2 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—
Runs On
What it takes to serve PhysicalEye-Decide-35B (34.6B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 69.2 GB | 83.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x MI325X $2.00 · 1x MI355X $2.59 |
| 8-bit | 34.6 GB | 41.5 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 17.3 GB | 20.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.
PhysicalEye-Decide-35B on every accelerator the SAVRN Index prices, at every precision
Model Card
By Physical AI, published under apache-2.0, revision 0c78df404606.
One forward pass. One calibrated decision. PhysicalEye-Decide-35B answers a question about an image by choosing among the options you give it, and returns a probability for every option. It does not write text. It reads the image, the question and the options once, and outputs a decision. Use it for visual inspection, yes/no checks, option selection and scoring on production lines, in apps and in robots. - A small decision readout sits on top of the backbone and scores each option directly. One request carries up to 4 images, an optional context (state) and one or more questions. Three question types are supported: choice (pick one option), noul (yes or no) and score (a value on a scale).…
Read Physical AI's full model card
One forward pass. One calibrated decision.
PhysicalEye-Decide-35B answers a question about an image by choosing among the options you give it, and returns a probability for every option. It does not write text. It reads the image, the question and the options once, and outputs a decision.
Use it for visual inspection, yes/no checks, option selection and scoring on production lines, in apps and in robots.
Results
| Test | Result |
|---|---|
| Egg inspection, held-out test (100 photos, good vs bad) | 98.0% accuracy, mean confidence 0.964 |
| AI2D diagram questions (200, 4 options) | 87.0% (chance 25%) |
| Latency per decision, median (one B200 GPU) | 0.051 s (p80 0.068 s) |
How it works
- Backbone: a vision-language Mixture-of-Experts model (35B total, about 3B active per token).
- A small decision readout sits on top of the backbone and scores each option directly.
- Output: one probability per option. Probabilities are temperature-calibrated.
Input format
One request carries up to 4 images, an optional context (state) and one or more questions.
Three question types are supported: choice (pick one option), noul (yes or no) and score (a value on a scale).
{
"images": ["<base64 image>"],
"state": "Egg inspection line, candling light.",
"questions": {
"q1": {"type": "choice", "question": "Is this egg good or bad?", "options": ["good", "bad"]}
}
}
The response has one entry per question key in answers, with the chosen option and its probability.
Files
model-*.safetensors: backbone weightsreadout.safetensors: decision readoutdecision_config.json: option codes, token ids and calibration temperature
Base model and license
- Built on FINAL-Bench/Darwin-35B-A3B-Opus, a Qwen3.5 MoE based model.
- Released under Apache-2.0. The upstream Qwen license notice applies to the base weights.
Training data credits
The decision readout was trained on openly licensed data:
- VQAv2 (CC BY 4.0, annotations), COCO captions and instances 2017 (CC BY 4.0, annotations; images keep their Flickr licenses)
- TextVQA 0.5.1 (CC BY 4.0), ScreenQA and RICO-SCA / Screen2Words on RICO (CC BY 4.0)
- CLEVR v1.0 (CC BY 4.0), A-OKVQA (Apache-2.0), VSR (CC BY 4.0), CORD v2 (CC BY 4.0)
- Open-Jev (CC0 1.0), MMLU auxiliary train (MIT), synthetic charts made in-house
- Good and Bad Eggs Identification Image Dataset, Md Mafiul Hasan Matin, Marium Jahan, Md. Saroar Jahan, Mendeley Data, doi:10.17632/mdty358x8m.1 (CC BY 4.0)
Configuration
- Architecture
- Qwen3_5MoeModel
- Context length (tokens)
- 262,144
- Layers
- 40
- Hidden size
- 2,048
- Attention heads
- 16
- Key/value heads
- 2
- Head dimension
- 256
- Vocabulary size
- 248,320
- Experts
- 256
- Experts active per token
- 8
- Model type
- qwen3_5_moe
Identity and Version
- Repository
- masterset-ai/PhysicalEye-Decide-35B
- Publisher
- Physical AI
- Task
- Image and text to text
- Modality
- Image and text
- Library
- transformers
- Parameters
- 34.6B parameters
- Languages
- moe
- Revision
- 0c78df4046067f2fd982320032f1c4183bfbe1f2
- First published
- 2026-10-10
- Last updated
- 2026-10-10
Files and Weights
25 files, 69.2 GB in total. The weights are 16 files totalling 69.2 GB in safetensors.
Weights16 files · 69.2 GB
Configuration4 files · 3.2 MB
Tokenizer2 files · 20.0 MB
Documentation1 file · 2.9 KB
Other1 file · 7.8 KB
Repository1 file · 1.6 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00015.safetensors | Weights | 5.0 GB | 956a56ca897f |
| model-00002-of-00015.safetensors | Weights | 4.5 GB | d31acbb53dc0 |
| model-00003-of-00015.safetensors | Weights | 5.0 GB | 29e3b1953831 |
| model-00004-of-00015.safetensors | Weights | 4.5 GB | 2a96a04f6e69 |
| model-00005-of-00015.safetensors | Weights | 4.5 GB | 98d2fcb5550b |
| model-00006-of-00015.safetensors | Weights | 4.5 GB | d1d8cf1643d9 |
| model-00007-of-00015.safetensors | Weights | 5.0 GB | 0143ea12d821 |
| model-00008-of-00015.safetensors | Weights | 4.5 GB | 8e8d5bfc02db |
| model-00009-of-00015.safetensors | Weights | 4.5 GB | d36dfde513be |
| model-00010-of-00015.safetensors | Weights | 5.0 GB | c3294e16b93f |
| model-00011-of-00015.safetensors | Weights | 4.6 GB | 4536d9f228f9 |
| model-00012-of-00015.safetensors | Weights | 5.0 GB | 6ffc66ef2bb5 |
| model-00013-of-00015.safetensors | Weights | 5.0 GB | 586753d03d00 |
| model-00014-of-00015.safetensors | Weights | 4.6 GB | 2fb387ab9479 |
| model-00015-of-00015.safetensors | Weights | 3.1 GB | 2bc1df47c958 |
| readout.safetensors | Weights | 1.0 MB | 418ee3d78bcc |
| config.json | Configuration | 3.3 KB | — |
| decision_config.json | Configuration | 4.4 KB | — |
| model.safetensors.index.json | Configuration | 3.1 MB | — |
| processor_config.json | Configuration | 1.4 KB | — |
| README.md | Documentation | 2.9 KB | — |
| chat_template.jinja | Other | 7.8 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| tokenizer.json | Tokenizer | 20.0 MB | 6f32ce20dc35 |
| tokenizer_config.json | Tokenizer | 1.3 KB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 69.2 GB
Download from Physical AI
Released by Physical AI through its official repository on Hugging Face. Read the license.
Built From
- Derived from FINAL-Bench/Darwin-35B-A3B-Opus
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 69.2 GB |
| 16-bit | 69.2 GB |
| 8-bit | 34.6 GB |
| 4-bit | 17.3 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About PhysicalEye-Decide-35B
How much GPU memory does PhysicalEye-Decide-35B need?
About 83 GB at 16-bit and 20.8 GB at 4-bit: the weights (34.6B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run PhysicalEye-Decide-35B on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use PhysicalEye-Decide-35B commercially?
Yes. PhysicalEye-Decide-35B is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
What is PhysicalEye-Decide-35B's context length?
262,144 tokens, from the maximum position embeddings in its published configuration.
Similar Models
Blackfrost Pro Series is a line of specialized models packaged as professional tools and workflows. Each release is designed around a defined operational role instead of open-ended chat. 1. a domain-tuned model; 2. a deliberate operating boundary; 3. a harness or integration surface; 4. deployment guidance; 5. a validation and provenance record. This structure is intended to extend across model engineering, marketing operations, logistics, and other bounded professional workflows while preserving a recognizable Blackfrost product and documentation standard. Ablagent-35B-A3B is the Blackfrost Pro Series toolkit for Abliteration and Domain Weight Modification workflows. It helps an operator…
MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 35B-A3B variant, with 35B-A3B being the first MoE model in the MiniCPM-V series. It builds on the family's single-image, multi-image, and video understanding capabilities, combining stronger foundation capabilities with efficient visual encoding and on-device inference. Notable features of MiniCPM-V 4.7 include: - The First Open-Source MoE Model with Lossless, Smooth On-Device Inference. With our Lookahead Expert Scheduling (LES) technique introduced alongside MiniCPM-V 4.7, combined with fine-grained layer-wise quantization and DSpark speculative sampling integration, MiniCPM-V 4.7 35B-A3B becomes the first…
In addition to the original formula, we have further enhanced Qwen2.5-VL-32B's mathematical and problem-solving abilities through reinforcement learning. This has also significantly improved the model's subjective user experience, with response styles adjusted to better align with human preferences. Particularly for objective queries such as mathematics, logical reasoning, and knowledge-based Q&A, the level of detail in responses and the clarity of formatting have been noticeably enhanced. In the past five months since Qwen2-VL’s release, numerous developers have built new models on the Qwen2-VL vision-language models, providing us with valuable feedback. During this period, we focused on…
In addition to the original formula, we have further enhanced Qwen2.5-VL-32B's mathematical and problem-solving abilities through reinforcement learning. This has also significantly improved the model's subjective user experience, with response styles adjusted to better align with human preferences. Particularly for objective queries such as mathematics, logical reasoning, and knowledge-based Q&A, the level of detail in responses and the clarity of formatting have been noticeably enhanced. In the past five months since Qwen2-VL’s release, numerous developers have built new models on the Qwen2-VL vision-language models, providing us with valuable feedback. During this period, we focused on…