Veylum

ARC Models / VEYLUM

The right light for your machine.

Five expressions of ARC

Start with the model that fits.

Fastest

ARC Ray

Quick answers, summaries, and drafts on modest hardware.

Installable · measured on one Windows host
Model and hardware details
Underlying weights
Qwen3.5-0.8B · Qwen team, Alibaba Group
GGUF conversion
unsloth/Qwen3.5-0.8B-GGUF
License
Apache License 2.0
Windows model download
507.85 MiB · Q4_K_M
Estimated system RAM
3.00 GiB minimum; 4.00 GiB recommended
Estimated free storage
1.17 GiB
Evidence
Loaded and evaluated on one Windows development machine. Results vary by hardware and context.

Download and model management happen in the desktop Models page. macOS runtime support remains unverified. Image input remains unavailable, including models whose upstream architecture supports images.

Efficient

ARC Prism

Everyday writing and instructions with a practical footprint.

Installable · measured on one Windows host
Model and hardware details
Underlying weights
Qwen3.5-4B · Qwen team, Alibaba Group
GGUF conversion
unsloth/Qwen3.5-4B-GGUF
License
Apache License 2.0
Windows model download
2.55 GiB · Q4_K_M
Estimated system RAM
6.00 GiB minimum; 8.00 GiB recommended
Estimated free storage
5.66 GiB
Evidence
Loaded and evaluated on one Windows development machine. Results vary by hardware and context.

Download and model management happen in the desktop Models page. macOS runtime support remains unverified. Image input remains unavailable, including models whose upstream architecture supports images.

Balanced

ARC Photon

Balanced reasoning and coding; hardware matters for this larger tier.

Verified here · reduced CPU mode
Model and hardware details
Underlying weights
Qwen3.5-9B · Qwen team, Alibaba Group
GGUF conversion
unsloth/Qwen3.5-9B-GGUF
License
Apache License 2.0
Windows model download
5.29 GiB · Q4_K_M
Estimated system RAM
10.00 GiB minimum; 16.00 GiB recommended
Estimated free storage
11.13 GiB
Evidence
Loaded and evaluated on one Windows development machine. Results vary by hardware and context.

On the measured 15.7 GiB machine with an 8 GiB framebuffer, Photon fell back to CPU at a reduced 2,048-token context. Larger hardware recommendations remain estimates.

Download and model management happen in the desktop Models page. macOS runtime support remains unverified. Image input remains unavailable, including models whose upstream architecture supports images.

Advanced

ARC Laser

A selected advanced tier for difficult reasoning and software work.

Selected · not yet installable
Model and hardware details
Underlying weights
Qwen3.5-27B · Qwen team, Alibaba Group
GGUF conversion
unsloth/Qwen3.5-27B-GGUF
License
Apache License 2.0
Windows model download
15.59 GiB · Q4_K_M
Estimated system RAM
24.00 GiB minimum; 32.00 GiB recommended
Estimated free storage
32.71 GiB
Evidence
Weights selected and pinned. Not downloaded, loaded, or evaluated; not offered for installation.

Download and model management happen in the desktop Models page. macOS runtime support remains unverified. Image input remains unavailable, including models whose upstream architecture supports images.

Workstation

ARC Spectra

A selected mixture-of-experts tier for larger workstations.

Selected · not yet installable
Model and hardware details
Underlying weights
Qwen3.5-35B-A3B · Qwen team, Alibaba Group
GGUF conversion
unsloth/Qwen3.5-35B-A3B-GGUF
License
Apache License 2.0
Windows model download
20.50 GiB · Q4_K_M
Estimated system RAM
32.00 GiB minimum; 64.00 GiB recommended
Estimated free storage
42.97 GiB
Evidence
Weights selected and pinned. Not downloaded, loaded, or evaluated; not offered for installation.

Download and model management happen in the desktop Models page. macOS runtime support remains unverified. Image input remains unavailable, including models whose upstream architecture supports images.