llama.cpp now runs Reflection's Beam, and its smallest GGUF file is 103.1 GB

llama.cpp, the open-source runtime that runs language models on ordinary computers, supports Beam as of build b11701. Reflection, an American lab, released the model, which has 501 billion parameters, on October 16, 2026, and llama.cpp added it the same day. Until then the documented options were Transformers, vLLM, and SGLang, and Hugging Face's guide sized the hardware in data-center GPUs.

What llama.cpp added

The change was merged into the main project on October 16, 2026. It has three parts: a script that converts Reflection's weights into a GGUF file, the model's computation graph, and a parser for the format Beam uses to call tools. In the same discussion, the author pointed to Unsloth's ready-made GGUF files.

Beam is a mixture-of-experts model. Each of its 49 expert layers holds 256 experts, which are smaller subnetworks, and a router picks eight of them for every token. Three dense layers, where all parameters are used, come first. Attention, the part of the model that decides which earlier tokens to weigh, alternates between layers that see only the last 512 tokens and layers that see the whole context without positional encoding.

The tool-call parser matters for agent work. Beam writes tool calls in its own variant of the Harmony format, and without the parser llama-server could not turn them into structured calls. The author reports that multi-turn sessions with parallel tool calls passed end to end through llama-server.

The author ran the real weights only on NVIDIA's CUDA platform. The other checks, including the architecture test on the CPU, used a tiny model with random weights and Beam's layer pattern. They confirm that the computation matches the reference implementation, but they say nothing about how fast the full model runs.

How large each file is

Quantization stores weights with fewer bits per value, which makes the file smaller at some cost in accuracy. Unsloth's repository has 11 quantized variants and a full copy in BF16. The sizes below are the sum of each file's parts as Hugging Face lists them.

• UD-IQ1_S is the smallest file, at 103.1 GB.
• UD-Q2_K_XL takes 186.9 GB.
• UD-Q3_K_XL takes 242.3 GB.
• UD-Q4_K_XL takes 306.5 GB.
• Q8_0 takes 532.8 GB.
• The unquantized BF16 copy takes about 1 TB.

What fits on which machine

A model runs at a usable speed only when the whole file, plus room for the context, fits in memory. This suggests that a machine with 128 GB can hold only the smallest file. At 256 GB, UD-Q2_K_XL fits, while UD-Q4_K_XL needs more than that. For comparison, Hugging Face's guide recommends four B200 GPUs, 720 GB in total, for the NVFP4 build of about 280 GB.

Fewer bits mean a larger departure from the original model. Unsloth publishes a chart of that departure for each file, measured as Kullback-Leibler divergence. No independent quality test of quantized Beam has been published.

No public speed figure for Beam in llama.cpp exists yet either. Only about 23.4 billion parameters are active per token. This suggests that memory capacity, not computing power, is the first limit for anyone who wants to run it.