Reflection Beam — local AI model by Reflection AI
Written by Jakub Rusinowski · Last updated
Reflection AI's first open-weight model, for coding, reasoning and agents: a sparse MoE released in BF16, FP8 and NVFP4 under Apache-2.0. Server-class to self-host.
Variants
The smallest Reflection Beam variant needs about 303 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache.
| Model | VRAM at Q4 | VRAM | Context | Run it |
|---|---|---|---|---|
| Beam 1.0 501B-A23B → 501B (23B active) | ~303.5 GB | 1,000,000 | cloud-hosted tag |
Memory is quantized weights plus overhead at Q4_K_M, from the same engine as the GPU & VRAM checker.
How to run Reflection Beam locally
Install Ollama, then pull the tag.
No local Ollama tag. Ollama lists Beam only as a cloud-hosted 'beam:cloud' tag, which runs on Ollama's servers, not your machine.
Pick a size above for its own VRAM figure, speed estimate and install command.
Licence
Commercial use permitted. No usage restrictions beyond attribution.
Applies to: Beam 1.0 501B-A23BRecommended GPU
The cheapest catalogued GPU that runs Reflection Beam locally (min 303 GB VRAM) is the Apple M3 Ultra (512 GB).
Reflection Beam — frequently asked questions
How much VRAM does Reflection Beam need?
Reflection Beam needs about 303 GB VRAM at Q4_K_M quantization for its smallest variant. Variants: Beam 1.0 501B-A23B (303 GB, Q4_K_M). On Apple Silicon, unified memory counts toward this requirement.
Can I run Reflection Beam on an RTX 4090 (24 GB)?
Reflection Beam's smallest variant needs about 303 GB, which exceeds a single RTX 4090 (24 GB). Use multiple GPUs, a higher-VRAM card, or Apple Silicon with large unified memory.
What quantization should I use for Reflection Beam?
Q4_K_M is the best balance of quality and VRAM for Reflection Beam in most cases. Choose Q8_0 for near-lossless quality if you have spare VRAM, or smaller quants (Q3/Q2) only when memory is tight.
How do I run Reflection Beam with Ollama?
Reflection Beam has no local Ollama tag — the published tag is cloud-hosted, so running it sends your prompts to a hosted GPU rather than your own machine.