Beam-1.0 on 4× NVIDIA RTX PRO 6000 Blackwell Server Edition: NVFP4 or FP8 on H200 NVL

Beam-1.0 hardware requirements on four Server Edition cards

Reflection's Beam-1.0 fits a server with four NVIDIA RTX PRO 6000 Blackwell Server Edition cards in its NVFP4 checkpoint of 293.6 GB, with 58.7 GiB left for the cache by our users-per-card rule: about 97 concurrent conversations of 32,768 tokens with an FP8 cache. In FP8 the model needs four H200 NVL, where 24.7 GiB remain for about 40 such conversations, or eight Server Edition cards. On memory and card count, NVFP4 on four Blackwell cards is the smaller build; the open point is software, because vLLM's recipe has no RTX PRO 6000 profile yet.

Reflection published the weights in October 2026 under Apache 2.0; vLLM's post on day-0 support is dated 16 October 2026. Beam is a text-only mixture-of-experts model with 256 experts, of which eight are routed per token, and a 1M-token context. This page takes the four-card decision for this one model. Our general page on the server with four RTX PRO 6000 Server Edition cards covers chassis and other models, and the card comparison of 4× RTX PRO 6000 or 4× H200 NVL the hardware without this model.

Our estimate by the users-per-card rule (83.04 GiB per Server Edition, 123.36 GiB per H200 NVL), FP8 KV cache of 0.605 GiB per conversation of 32,768 tokens at full length. Checkpoint sizes: sums of the safetensors files in reflection/Beam-1.0-501B-A23B-NVFP4 and -FP8 on Hugging Face, read on 24 October 2026.

Three checkpoints and what each one stores

The three repositories hold the same instruction-tuned model. The BF16 repository lists 188 shards and about 1 TB; vLLM's recipe gives 1,003 GB on disk. The FP8 checkpoint was made with llm-compressor, its card states, and the NVFP4 one with NVIDIA Model Optimizer by post-training quantisation.

Hugging Face file lists of the three reflection repositories, summed on 24 October 2026; BF16 size from vLLM's Beam recipe of 16 October 2026; GiB conversion by our arithmetic.

The NVFP4 file hf_quant_config.json shows what is in 4 bits. Model Optimizer quantised the expert weights in blocks of 16 values and left out the embeddings, the output layer, the first three layers, which are dense, and in every MoE layer the attention and the router. The file leaves the KV cache unquantised (kv_cache_quant_algo is null); vLLM's recipe sets an FP8 cache with --kv-cache-dtype fp8. The NVFP4 card lists its scores against BF16: 80.7 against 81.6 per cent on SWE-bench Verified and 89.3 against 90.6 on GPQA Diamond, with temperature 0.7.

The file sizes also show the cost per layer. Each dense layer is a file of 839 MB, and each MoE layer one of 5.88 GB. With the 629 MB of a dense layer's MLP from config.json, the attention takes about 210 MB per layer in BF16, and one expert about 22 MB in NVFP4, which matches 39.3M parameters at 0.5625 bytes.

How much memory one conversation needs

Beam keeps most of its cache short. Its config.json defines 52 layers: 43 use sliding-window attention over 512 tokens and nine keep full attention, each with 8 KV heads of 128 values. vLLM's day-0 post counts "9 × 2 × 8 × 128 values per token, or 36 KiB in BF16 and 18 KiB in FP8". The sliding layers add a fixed amount per conversation, 43 MiB in FP8 by our arithmetic.

Our estimate from config.json (52 layers, 9 full attention, 8 KV heads, head dim 128, window 512) and vLLM's post of 16 October 2026; conversations at full length in the 58.7 GiB left on four Server Edition cards.

With tensor parallelism over four cards, each card holds two of the eight KV heads, so the cache divides evenly; vLLM's recipe states about 73 GB of weights per GPU for NVFP4 at TP=4. These counts are conversations in flight at the same moment; our guide to private ChatGPT server sizing by company size turns headcount into that number. The rule itself is derived in how many users one RTX PRO 6000 can serve. Speculative decoding with Reflection's DFlash2 drafter costs "about 1.7 GB per GPU", the recipe says, which lowers the 32K figure on four cards to about 86.

NVFP4 on Blackwell against FP8 on Hopper

The RTX PRO 6000 computes FP4 on its Tensor Cores. vLLM's recipe states that on Blackwell it runs Beam's NVFP4 experts as W4A4, while on Hopper "vLLM serves the FP4 experts through Marlin with BF16 activations", a weight-only path that saves memory and computes at 16-bit speed. Its Hopper profile for NVFP4 is eight H100 at TP=8; for four H200 NVL in NVFP4 the table above gives a memory fit only. The recipe's FP8 profile is eight H200 at TP=8, and it names the minimum GPU memory per checkpoint: 353 GB for NVFP4, 604 GB for FP8 and 1,204 GB for BF16.

Four Server Edition cards have 384 GB, above the 353 GB minimum for NVFP4. Four H200 NVL have 564 GB, below the 604 GB the recipe asks for FP8, while our rule leaves 24.7 GiB, about 40 conversations of 32K. That build works for a pilot or a small team, and it is tight for long contexts. Eight Server Edition cards in FP8 meet the recipe's figure and keep 195.6 GiB. The formats themselves are explained in our guide to FP8, NVFP4, MXFP4 and INT4, and FP4 models on H200 NVL and Ada GPUs covers the weight-only path.

If quality must match BF16 as closely as possible, the FP8 build is the safer choice; the NVFP4 card's own table shows a loss of about one point on most of its benchmarks. We supply both four-card servers and the eight-card one; send us your context length and the peak number of requests, and our team replies with the configuration and the figures for each checkpoint.

Software status on the RTX PRO 6000 (SM120)

As of 24 October 2026, Beam is not in a vLLM release. The recipe, updated 16 October 2026, names vLLM 0.33.0 as its minimum and states that "Beam support has not shipped in a vLLM release yet", so it serves from the vllm/vllm-openai:reflection-beam image. The latest release on vLLM's GitHub page is 0.32.0 of 18 October 2026. The model card gives no hardware list and no commands.

The recipe's profiles are four B200 or B300 or one GB200 or GB300 tray for NVFP4 at TP=4, eight H100 for NVFP4 through Marlin, eight H200 for FP8 and eight B200 or B300 for BF16. The RTX PRO 6000 is compute capability 12.0, a different target from the B200, and the recipe does not mention it. The general SM120 status is better than in the spring: vLLM 0.30.0 made native W4A4 NVFP4 kernels the default on SM120, as our guide to SM120 software support records. Whether the Beam image includes them, and which attention backend it picks for gated attention with a sliding window, has to be read from the start-up log on the delivered server.

Plan the first start. The recipe's NVFP4 command sets VLLM_ENGINE_READY_TIMEOUT_S=3600 because loading takes about 10 minutes, and the NVFP4 command runs --kv-cache-dtype fp8 with Beam's own tool-call and reasoning parsers.

TP=4 or expert parallelism over PCIe

The recipe runs NVFP4 with --tensor-parallel-size 4. On four Server Edition cards this all-reduce crosses PCIe, because the card has no NVLink. By default vLLM disables its custom all-reduce on more than two PCIe-only GPUs and uses NCCL; release 0.30.0 (22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink, which our article on RTX PRO 6000 without NVLink covers. The recipe also has a data-parallel layout with expert parallelism and its MegaMoE backend; the recipe allows it only as data plus expert parallelism, and vLLM's post measured it on four GB300, so we do not count on it for these cards.

For speed, the file sizes give an upper bound. By our estimate one token reads the three dense layers (2.5 GB), in each of the 49 MoE layers the attention plus eight experts (about 0.39 GB) and the output layer (1.6 GB), about 23 GB in total. Four cards read 4 × 1,597 GB/s, so one stream cannot exceed about 275 tokens per second, an upper bound by our estimate before any communication over PCIe. Batched requests hit more experts per step, so the bound does not scale with users. vLLM's day-0 post reports 708 tokens per second for one request with DFlash2 at TP=8; its optimisation runs were on GB200.

Power, host memory and storage

The four-card server draws what any four-card Server Edition build draws. With the cards at 600 W and two AMD EPYC 9355 at 280 W we estimate about 3.4 kW at the wall, or about 2.7 kW with the cards capped at 450 W, the canon of our four-card page. At 600 W that is above the 2.94 kW we plan per 230 V 16 A feed, so the rack position needs 32 A or three-phase; our guide to 16 A, 32 A or three-phase for GPU servers has the table. Four H200 NVL at 600 W come to the same estimate. A measurement on the delivered system confirms or replaces these figures.

NVIDIA's certified-systems guide asks for host memory of at least twice the GPU memory, so we quote 768 GB of DDR5 ECC, as in the inference node on our AI servers page. Keep the NVFP4 checkpoint and the FP8 one side by side if you want to compare them on the same server: 797 GB together, more if the BF16 weights are stored for fine-tuning. Our guide to NVMe storage for an LLM server sizes the drives.

Which build fits which plan

• A department or pilot with long tasks: four Server Edition cards in NVFP4, about 97 conversations of 32K, after a start-up test of the reflection-beam image on that card.
• Four H200 NVL: in FP8 about 40 conversations of 32K, below the recipe's 604 GB minimum; the NVFP4 checkpoint run weight-only would leave far more room, but no recipe profile documents it on four cards.
• FP8 quality and room to grow: eight Server Edition cards, 195.6 GiB left in FP8, or NVFP4 with room for several hundred conversations. Our page on eight RTX PRO 6000 Server Edition for large LLMs covers the chassis.
• Several models on one server: Beam in NVFP4 leaves 14.7 GiB per card; a second model belongs on other cards.

Other models in the same size class on the same four cards, GLM-5.3-Flash, DeepSeek-V4-Flash, Qwen3.8-Flash-Next and MiMo-V2.6-Flash, are compared in 300B MoE models on four RTX PRO 6000. GLM-5.3-Flash has its own page, GLM-5.3-Flash on 4× RTX PRO 6000 Server Edition.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with four or eight RTX PRO 6000 Server Edition cards or four H200 NVL, assembled and tested before delivery, on one EU contract and invoice, with manufacturer warranty on every component handled through us. Quantities and lead time are confirmed in writing with the quote; the ordering steps are on our page to buy a 4× RTX PRO 6000 Server Edition server. Setting up Beam, retrieval and the serving stack is part of our Private AI/ML service; describe your use of Beam-1.0 and we say which build and checkpoint we would start from.

FAQ