Beam 1.0 VRAM requirements at Q4 and 1M context
Beam 1.0 needs about 351 GB in this Q4_K_M, 1,048,576 token, batch one scenario. The Q4 weights take 281 GB of that and the KV cache takes 38.7 GB.
Beam 1.0 is Reflection's open weight sparse mixture of experts model, with 501B total parameters and 23B active for each token. The 23B figure describes compute. Memory follows the 501B. The router picks 8 of the 256 experts per token and can pick different ones on the next token, so every expert has to stay resident if you want full speed. This page budgets all of them at Q4_K_M, which the calculator treats as 0.56 bytes per parameter, then adds an FP16 KV cache for one sequence and a 10% runtime allowance. That gives 351.2 GB. It's a planning figure, not a measured footprint, and the next sections show which parts of it move when you change the context.
By TechCompare · Updated
Calculator
Estimated VRAM required
351 GB
501B params at Q4_K_M, 1,048,576 token context, batch 1, inference.
Estimate accuracy: Planning estimate. Rounded model counts, tensor formats, cache packing, and workspace can change actual allocation. Confirm the exact checkpoint and serving engine with a load test.
Beam 1.0 has 9 full attention layers and 43 layers with a 512 token sliding window, all with 8 KV heads at head dim 128. Only the full layers grow with context. All-resident weights include the full 501B pool.
Hardware that fits
Apple entries show installed unified memory separately from a model budget based on an 8 GB reserve. That budget is a planning heuristic, not a fixed hardware limit.
No single GPU in our catalog has enough memory. Multi-GPU or CPU offload required.
How this is calculated
The architecture is what makes this page different from most 1M context models. Beam has 52 layers with hidden size 5120, and every attention layer uses 8 KV heads at head dimension 128. Only 9 of the layers use full attention. That's every sixth layer starting at layer 5, plus layer 51. The other 43 look back over a 512 token window and nothing more. Those 43 layers hold a fixed amount of cache no matter how long the session runs. Only the 9 full layers grow, at 36,950 bytes per token in FP16. At 1,048,576 tokens that adds up to 38.7 GB, and at 262,144 tokens it's 9.8 GB. At 32K it's 1.3 GB, and at 8K just 0.39 GB, mostly the window layers' fixed buffers. The totals follow: 319 GB at 256K, 310 GB at 32K and 351 GB at the full 1M. The cache is the smaller part of the budget at every context this model supports. Cutting context helps, but it only trims the edge of a budget that the weights set. Batch changes that picture, because each extra sequence needs its own cache while the weights are shared. Four sessions at 256K come to 352 GB in this estimate. The estimate leaves out prompt processing buffers, communication workspace for splitting the model across cards, and any framework allocations beyond the 10% allowance, so measure peak memory on your own stack at the context you plan to serve.
Verdict
Plan Beam 1.0 as a multi card or multi node deployment and treat the context length as a small adjustment. At the full 1M scenario the estimate takes 4 cards of a nominal 96 GB each, and at 32K it still takes 310 GB. Eight cards of 96 GB give 768 GB in total, which clears the full context figure on paper. Four such cards hold the short context case on paper but not the million token one. Uneven sharding, per card runtime overhead and the serving engine's support for this architecture all decide whether a paper fit becomes a working setup. The weights come in BF16, FP8 and NVFP4 releases, so pick a release first and compare its file size with the 281 GB used here before buying hardware.
More Beam scenarios
Related guides
Frequently asked questions
How much VRAM does Beam 1.0 need at Q4?
Beam 1.0 only activates 23B parameters, so why plan for 501B?
Why is the 1M context cache so small for a model this size?
How many 96 GB GPUs does Beam 1.0 take?
Should I use the FP8 or NVFP4 release instead of a Q4 conversion?
What does this estimate leave out?
Related tools
RAM Latency Calculator
Convert DDR3/DDR4/DDR5 timings (CL, tRCD, tRP, tRAS) into true latency in nanoseconds.
Use tool ➜Power Cost Estimator
Estimate annual electricity costs for your PC, Server, or TV.
Use tool ➜Data Transfer Calculator
Estimate transfer times for files over USB, WiFi, Ethernet, and more.
Use tool ➜Memory and Storage Throughput Visualizer
Visualize the massive speed difference between CPU cache, RAM, and storage.
Use tool ➜