Can I Run Beam 1.0 501B Locally? (2026): Yes, on 256GB
Reflection put Beam-1.0-501B-A23B on Hugging Face on 7 October 2026. It is a 501B mixture-of-experts model with 23.4B active parameters, a 1M-token context and an Apache 2.0 licence. We summed every GGUF file in bytes from the Hugging Face API, did the cache arithmetic from config.json, and checked llama.cpp, vLLM, SGLang and MLX. The short answer: 256GB of memory runs a good 3-bit file. 128GB runs only experimental 1-bit builds.
Bottom Line
- Yes, with 256GB of memory. A 256GB Mac Studio M5 Ultra holds the 3-bit UD-IQ3_XXS (183.61 GiB). AtomicChat measured it at 96.13% top-1 agreement with BF16.
- The cache is small. Only 9 of 52 layers keep a full cache: 36,864 bytes per token. The full 1M-token context costs 36.0 GiB at f16.
- 128GB machines run only experiments. One DGX Spark runs a 1.5-bit VQ build at ~25 tok/s in a custom engine (community-reported). Standard GGUFs need 1 bit.
- A GPU plus a lot of RAM works. A 24-32GB card with 192GB of RAM holds UD-IQ3_XXS with the experts in RAM.
- Two RTX PRO 6000 cards are the fast route. A 3-bit EXL3 build decodes at 117 tok/s with 256GB of host RAM (community-reported).
- Three DGX Sparks run the official NVFP4 weights at ~18 tok/s (community-reported). Two Sparks are too small for it.
- Skip the 255B REAP prune unless you only code. It keeps HumanEval but drops MMLU-Pro from 0.843 to 0.702.
Ready to buy? See the tested hardware list with current prices.
What Beam 1.0 501B Is
Figures come from the model card, its config.json and the Hugging Face API, read on 26 October 2026.
| Spec | Value |
|---|---|
| Repo | reflection/Beam-1.0-501B-A23B |
| Created on Hugging Face | 7 October 2026 |
| Parameters in the checkpoint | 501,267,612,352 (BF16 plus 64M F32) |
| Active per token | 23.4B (card); our count from config.json is 23.43B |
| Layers | 52: 3 dense + 49 MoE |
| Attention | Gated GQA: 48 query heads, 8 KV heads, head size 128 |
| Layer pattern | 43 sliding-window (512 tokens) + 9 full attention |
| Experts | 256 routed per MoE layer, 8 active, no shared expert |
| Context | 1,048,576 tokens, native |
| Inputs | Text only |
| Licence | Apache 2.0 |
| Official checkpoints | BF16 (933.81 GiB), FP8 (468.72 GiB), NVFP4 (273.47 GiB) |
| Drafter | DFlash2, a separate 3.59 GiB repo for speculative decoding |
Our count matches the card. The always-read weights are 7.24B: attention with its output gate, the 3 dense layers, the routers and the output head. The 8 active experts add 15.41B. The token table adds 0.78B. The total is 23.43B.
What the card claims. These are Reflection’s own numbers, not ours:
| Benchmark (card) | BF16 | NVFP4 |
|---|---|---|
| SWE-bench Verified | 81.6% | 80.7% |
| Terminal Bench v2.1 | 80.1% | 78.4% |
| GPQA Diamond | 90.6% | 89.3% |
| LiveCodeBench | 88.3% | 88.3% |
We did not run any benchmark. One discussion on the official repo is titled “Reward-hacking garbage”. Another user calls it an excellent coding executor. Test it on your own tasks.
Every GGUF Size, in Bytes
We summed every file in the Hugging Face file trees on 26 October 2026. GiB = bytes / 1,073,741,824. Split files are added together.
| File | Repo | Bytes | GiB |
|---|---|---|---|
| VQ (about 1.5 bits, own engine) | Baekpica | 102,024,335,169 | 95.02 |
| UD-IQ1_S | unsloth | 103,059,961,380 | 95.98 |
| IQ1_S | bartowski | 109,586,471,392 | 102.06 |
| UD-IQ1_M | unsloth | 123,612,050,980 | 115.12 |
| AD-IQ2_XXS | AtomicChat | 133,789,612,416 | 124.60 |
| UD-IQ2_XXS | unsloth | 133,885,310,500 | 124.69 |
| Lite-Beam IQ2_XXS | tablinum | 144,421,041,184 | 134.50 |
| IQ2_XS | bartowski | 151,720,589,920 | 141.30 |
| AD-IQ2_M | AtomicChat | 164,814,354,944 | 153.50 |
| UD-Q2_K_XL | unsloth | 186,877,720,196 | 174.04 |
| AD-IQ3_XXS | AtomicChat | 197,111,315,008 | 183.57 |
| UD-IQ3_XXS | unsloth | 197,154,584,164 | 183.61 |
| UD-Q3_K_XL | unsloth | 242,308,069,188 | 225.67 |
| IQ4_XS | bartowski | 273,062,460,480 | 254.31 |
| UD-Q4_K_XL | unsloth | 306,533,349,412 | 285.48 |
| Q4_K_M | bartowski | 314,533,641,408 | 292.93 |
| UD-Q5_K_XL | unsloth | 369,474,123,812 | 344.10 |
| Q6_K | bartowski | 441,778,339,360 | 411.44 |
| Q8_0 | unsloth / bartowski | 532,794,735,998 | 496.20 |
| BF16 | unsloth | 1,002,671,882,528 | 933.81 |
DevQuasar also publishes Q2_K (169.56 GiB) to Q5_K_M (330.86 GiB). The Baekpica figure is the whole repo, including its 1.98 GiB drafter and its tools.
How much quality each step costs. AtomicChat measured each file against the BF16 original on A100s. “Top-1” is how often the quant picks the same next token:
| File | GiB | KL divergence | Top-1 agreement |
|---|---|---|---|
| AD-IQ2_XXS (AtomicChat) | 124.60 | 0.0674 | 92.82% |
| AD-IQ2_M (AtomicChat) | 153.50 | 0.0337 | 94.90% |
| UD-Q2_K_XL (unsloth) | 174.04 | 0.0209 | 95.93% |
| UD-IQ3_XXS (unsloth) | 183.61 | 0.0189 | 96.13% |
Bartowski measured his own files with a different text. His Q4_K_M scored KL 0.0250 and his IQ1_S scored 0.6917. That is about 28 times the drift of Q4_K_M. Treat every 1-bit file as a demo, not a daily driver.
The 2-bit to 3-bit step is the one that matters. KL divergence falls by more than two thirds from AD-IQ2_XXS to UD-Q2_K_XL. We found no matched measurement above 3 bits.
Why the Context Costs So Little
Only the 9 full-attention layers keep a cache that grows with context. The 43 sliding-window layers keep only the last 512 tokens.
1. The full-attention cache. Config.json gives num_key_value_heads 8 and head_dim 128. layer_types lists 9 full_attention layers.
f16: 2 (K and V) x 8 heads x 128 x 2 bytes x 9 layers = 36,864 bytes per token
32,768 tokens = 1.1 GiB
131,072 tokens = 4.5 GiB
262,144 tokens = 9.0 GiB
1,048,576 tokens = 36.0 GiB (q8_0 cache: about 19.1 GiB)AtomicChat’s card says “about 37 MB per 1,000 tokens at f16”. The OpensourceWTF card says 36,864 B per token. Both match.
2. The sliding-window rings. Each of the 43 layers holds a 512-token window plus a working batch. Baekpica’s budget puts this at 0.42 GiB. It does not grow with context.
3. The full budget. Weights + cache + 0.5 GiB rings + a compute buffer. We budget 2 GiB for the buffer. We did not measure it.
4. No RoPE scaling is involved. The 9 full layers use no positional encoding (NoPE). The 1M context is the model’s own setting, not a YaRN stretch.
Fit Table by Hardware
PC rows use VRAM plus system RAM, minus about 8 GiB for the OS. Every row is our arithmetic unless it says community-reported.
| Hardware | Usable memory | Best file | Context | Notes |
|---|---|---|---|---|
| 128GB Mac (M5 Max) | 107.52 GiB default limit | UD-IQ1_S (95.98) | 32K | 1-bit only; not worth it |
| DGX Spark | 117.2 GiB available (community) | Baekpica VQ (95.02) | 64K measured | ~25 tok/s, custom engine |
| Strix Halo 128GB | not checked for this model | UD-IQ1_S (95.98) | short | 1-bit; no report found |
| 24GB GPU + 128GB DDR5 | ~144 GiB | AD-IQ2_XXS (124.60) | 131K | Experts in RAM |
| 32GB GPU + 128GB DDR5 | ~152 GiB | IQ2_XS (141.30) | 64K | Experts in RAM |
| 2 x DGX Spark | ~234 GiB | UD-IQ3_XXS (183.61) | 131K | Needs llama.cpp RPC; no run found |
| 24-32GB GPU + 192GB DDR5 | ~208-216 GiB | UD-IQ3_XXS (183.61) | 262K | Experts in RAM |
| 256GB Mac Studio M5 Ultra | 240 GiB with raised limit | UD-IQ3_XXS (183.61) | 1M | 36 GiB cache fits |
| 2 x RTX PRO 6000 + 256GB RAM | 192GB VRAM | diffbot EXL3 3bpw (180.43) | 128K | 117 tok/s (community) |
| 3 x DGX Spark | 3 x 117.2 GiB | NVFP4, padded for TP3 (287.68) | 131K | ~18 tok/s (community) |
| 24-32GB GPU + 256GB DDR5 | ~272-280 GiB | IQ4_XS (254.31) | 131K | Experts in RAM |
| 4 x RTX PRO 6000 | 384GB VRAM | Official NVFP4 (273.47) | long | vLLM; 97.8 tok/s (community) |
A 256GB Mac Studio holds 3 bits and the full 1M context. UD-IQ3_XXS (183.61 GiB) + 36.0 GiB cache + 0.5 GiB rings + 2 GiB buffer is about 222 GiB. That is under the 240 GiB limit. Raise the limit first with sudo sysctl iogpu.wired_limit_mb=245760, the value the OpensourceWTF card uses on a 256GB M5 Ultra. UD-Q3_K_XL (225.67 GiB) also loads at 32K, but it leaves macOS very little room.
The 128GB Mac fits only 1 bit. A 128GB M5 Max reports a 115,448.73 MB working set in llama.cpp issue #21680. That is 107.52 GiB. UD-IQ1_S + 1.1 GiB cache + 2.5 GiB is about 99.6 GiB, so it loads. The next file, UD-IQ1_M (115.12 GiB), does not fit without a raised limit.
The DGX Spark route is real but experimental. Baekpica’s VQ build stores the experts at about 1.5 bits with vector quantization. It runs only in the ds4-dfm-rs CUDA engine, not in llama.cpp. Its card warns of reasoning loops and broken characters. It plans the full 1M context at 115.4 of 117.2 GiB with an FP8 cache, but no run past 64K was made.
Two Sparks are too small for the official NVFP4. The NVFP4 repo is 273.47 GiB. Two Sparks give about 234 GiB. Bespoke-ALI padded the weights with zeros so vLLM can split them across three Sparks. Their card says the padded build “does not load at TP=2, 4 or 8”.
The diffbot build needs more than two cards. It keeps the coldest 31 experts of each layer (12%) in pinned host RAM. Its card asks for about 256GB of host RAM and two 96GB cards with PCIe peer-to-peer. It ships its own kernels and a patched vLLM.
Is the 255B REAP Prune Worth It?
mrfakename’s Beam-1.0-REAP-255B-A23B removes half the experts. Each MoE layer keeps 128 of 256. Active parameters stay at about 23B, so it is not faster per token. It is only smaller.
| Original | REAP 255B | |
|---|---|---|
| Total parameters | 501.27B | 254.61B |
| BF16 size | 933.81 GiB | 474.31 GiB |
| NVFP4 size | 273.47 GiB (official) | 144.14 GiB (peterhdd) |
| Perplexity (mrfakename) | 3.32 | 3.70 |
| Top-1 agreement (mrfakename) | 100% | 84.7% |
peterhdd quantized it to NVFP4 and ran benchmarks on RTX PRO 6000 cards (community-reported, one run each):
| Benchmark | Official NVFP4 (4 GPUs) | REAP NVFP4 (2 GPUs) |
|---|---|---|
| MMLU-Pro (420 questions) | 0.843 | 0.702 |
| IFEval strict | 0.939 | 0.896 |
| HumanEval pass@1 | 0.963 | 0.951 |
Coding survives, knowledge does not. The experts were chosen on code and agent text. Health fell from 0.87 to 0.57, law from 0.77 to 0.53, history from 0.60 to 0.43 (30 questions each). The REAP build fits two RTX PRO 6000 cards entirely in VRAM and decodes at about 80 tok/s for one user. Pick it only for a coding-only server. For general use, the 3-bit full model at 183.61 GiB keeps every expert.
No REAP GGUF existed when we checked. The REAP build runs in vLLM only.
How Fast Will It Be?
These are owner measurements from model cards and threads (community-reported):
| Setup | Build | Prefill | Decode |
|---|---|---|---|
| 2 x RTX PRO 6000 Max-Q, PCIe, DFlash2 | diffbot EXL3 3bpw | 4,720-4,770 tok/s (46K prompt) | 110-122 tok/s |
| 4 x RTX PRO 6000 | Official NVFP4, vLLM | n/a | 97.8 tok/s |
| 2 x RTX PRO 6000 | REAP NVFP4A16, vLLM | n/a | 80.4 tok/s |
| Mac Studio M5 Ultra 256GB | OpensourceWTF NVFP4 MLX stream | 1,063 tok/s (16K prompt) | 29.1-29.5 tok/s |
| 1 x DGX Spark | Baekpica VQ, ds4-dfm-rs | 1,241 tok/s (2K-64K) | 25.23 tok/s |
| 3 x DGX Spark | Bespoke-ALI NVFP4 TP3, vLLM | 0.7 s TTFT at 1K | ~18 tok/s |
| 4 x Tesla P100 + system RAM | AtomicChat AD-IQ2_XXS, llama.cpp | n/a | ~4-5 tok/s |
The Mac pack streams the experts from the SSD into a cache on the GPU. Its card measured a 98.6% expert hit rate. Baekpica’s GB10 speeds come from the first build; the revised files were not re-measured.
We found no llama.cpp speed for a single GPU plus RAM. The rows below are bandwidth ceilings, not predictions. Real speed lands below them.
Weights read per token (our count from config.json):
8 routed experts x 3 x 5,120 x 2,560 x 49 layers = 15.41B (in RAM if offloaded)
attention + gate, dense layers, routers, head = 7.24B (on the GPU)
PC, experts in RAM, dual-channel DDR5-4800 = 76.8 GB/s:
UD-IQ2_XXS, ~2.14 bits/weight (file average): 15.41B x 2.14 / 8 = 4.12 GB -> ~19 tok/s ceiling
UD-IQ3_XXS, ~3.15 bits/weight (file average): 15.41B x 3.15 / 8 = 6.07 GB -> ~13 tok/s ceilingThe bits per weight are each file’s bytes x 8 / 501.27B. Experts that fit in spare VRAM raise the ceiling a little. Four DIMMs often run slower than two, which lowers it.
The always-read part is big. 7.24B weights are read on every token. Tablinum’s card lists Unsloth’s attention type as Q6_K, so plan on about 6 GB on the GPU. A 24GB card holds it with room for the cache.
Software Support, Checked 26 October 2026
| Runtime | Status | Source |
|---|---|---|
| llama.cpp | Supported from release b11701. PR #30670 merged 16 October 2026 | GitHub |
| llama.cpp build to use | b11804+ (bartowski’s files) | bartowski card |
| Known llama.cpp bug | Issue #30735: reply stuck in the thinking block (open) | GitHub |
| Ollama | Only beam:cloud. No local tag | ollama.com |
| LM Studio, koboldcpp, Jan | Listed by bartowski’s card. We did not test them | bartowski card |
| vLLM | PR #62074 merged 19 October, after v0.32.0 (18 October). Use a nightly or vllm/vllm-openai:reflection-beam | GitHub, peterhdd card |
| SGLang | Model PR #45256 merged on main 25 October, after v0.5.22 | GitHub |
| MLX | No Beam model in mlx-lm. The OpensourceWTF pack needs mlx-stream PR #20 and mlx-serve PR #896, both open | GitHub, pack card |
| ExLlamaV3 | diffbot’s EXL3 build runs through its own patched vLLM | diffbot card |
Run the 3-bit file on a 256GB machine with:
llama-server -hf unsloth/Beam-1.0-501B-A23B-GGUF:UD-IQ3_XXS -c 131072 --jinja \
--temp 0.7 --top-p 0.99 --repeat-penalty 1.0 --presence-penalty 0 --frequency-penalty 0 --dry-multiplier 0Turn the repetition samplers off. Unsloth’s reply on their GGUF thread says repetition penalties and DRY push Beam away from its own channel tags. The reply can then stay stuck in the thinking block. The reporter says these settings did not fix his case, so issue #30735 is still open.
On a PC with a 24-32GB card, add -ngl 99 --n-cpu-moe 52 and lower --n-cpu-moe until the GPU is full. Our —n-cpu-moe guide shows how. Beam always reasons first. Set the budget with --chat-template-kwargs '{"reasoning_effort": "medium"}', as AtomicChat’s card shows.
What to Buy
Prices as of October 2026, from our price reference. The read date is in brackets.
| Option | Price | What it runs |
|---|---|---|
| Mac Studio M5 Ultra 256GB (64-core GPU, 1TB) | $9,499 (apple.com, 22 October) | UD-IQ3_XXS with 1M context, or the MLX pack at ~29 tok/s |
| 3 x NVIDIA DGX Spark 128GB | $6,950 list each since 2 October 2026 (was $4,699); ASUS GX10 1TB $5,999 each (26 October) | Official NVFP4 at ~18 tok/s |
| 2 x RTX PRO 6000 96GB + 256GB host RAM | $14,999.99 per card (Central Computer, 30 September); RAM not priced | EXL3 3bpw at ~117 tok/s |
| DDR5 128GB (2 x 64GB) for a 24GB-card PC | from $2,099.99 (Newegg, 1 October) | AD-IQ2_XXS with the experts in RAM |
The 256GB Mac Studio is the simplest buy. It holds the 3-bit file and the full context in one box. Apple still sells the 256GB M5 Ultra; the 512GB tier was not orderable when our price reference last checked. We have no affiliate link for it. Configure it on apple.com.
For speed, two 96GB cards win. Two RTX PRO 6000 96GB cards run diffbot’s 3-bit build at about 117 tok/s. That is four times the Mac’s measured speed. The pair costs about three times as much, plus a workstation with 256GB of RAM.
Three Sparks are the CUDA middle route. Three NVIDIA DGX Spark 128GB boxes run Reflection’s own NVFP4 weights. Three at the $6,950 list cost $20,850, $5,850 more than one RTX PRO 6000 at $14,999.99. Three ASUS GX10 1TB boxes cost $17,997. They decode at about 18 tok/s and took 17 minutes to start in the owner’s test. One Spark alone runs only the experimental 1.5-bit build.
Already own a 24GB or 32GB card? Buy RAM, not a new machine. 128GB of DDR5 runs a 2-bit file. 192GB runs the 3-bit file. Expect single-digit to low-teens tok/s.
Do not buy a 128GB box for Beam. Every 128GB route is 1 to 1.5 bits. A 128GB machine is a good buy for 120B-class models such as Ornith 1.5 122B, not for this one.
Honest Caveats
- We did not run the model. Every fit row is arithmetic from file sizes and config.json.
- Every speed number is an owner’s post. Most come from one setup and a few runs.
- The 2 GiB compute buffer is a budget. Check the buffer sizes in your
llama-serverlog. - The benchmark table is Reflection’s. Community threads both praise and question it.
- The DGX Spark memory figure is second-hand. The 117.2 GiB comes from Baekpica’s
--check-configreading. - KL numbers come from different texts. AtomicChat’s and bartowski’s tables are not directly comparable with each other.
- The 2 x Spark row is untested. We found no published llama.cpp RPC run of Beam.
Sources
- Beam-1.0-501B-A23B model card: parameters, layers, attention type, context, licence, benchmark claims, numerics
- config.json: layer_types, mlp_layer_types, num_key_value_heads, head_dim, sliding_window, experts, moe_intermediate_size, max_position_embeddings
- Hugging Face API: creation date, parameter count; model.safetensors.index.json: tensor names (no shared expert)
- File trees for unsloth, bartowski, AtomicChat, Baekpica, tablinum, DevQuasar, and the official FP8, NVFP4 and DFlash2 repos: byte sizes
- AtomicChat GGUF card: KL divergence and top-1 agreement, cache per token, reasoning_effort flag
- bartowski GGUF card: build b11804, bits per weight, KL divergence for Q4_K_M and IQ1_S, compatible apps
- Baekpica VQ-GGUF card: DGX Spark speeds, 117.2 GiB available, 1M budget, warnings (community-reported)
- Bespoke-ALI TP3 card: 3 x Spark speeds and padding (community-reported)
- diffbot EXL3 card: 2 x RTX PRO 6000 speeds, host RAM need, KL divergence (community-reported)
- OpensourceWTF MLX stream card: M5 Ultra speeds, wired limit, open PRs (community-reported)
- mrfakename REAP card and peterhdd REAP NVFP4 card: pruning method, perplexity, MMLU-Pro, IFEval, HumanEval, speeds
- tablinum Lite-Beam card: 134.50 GiB 2-bit file, 156 GB DRAM reference layout
- Project-Apollo Beam result: 4 x P100 + RAM speed (community-reported)
- llama.cpp PR #30670, issue #30735 and issue #21680: Beam support, the thinking-block bug, the 128GB M5 Max working set
- vLLM PR #62074, vLLM releases, SGLang PR #45256: merge dates
- Unsloth GGUF discussion #1: sampler settings
- Ollama library: beam: cloud tag only
- OpenClaw DC hardware price reference, last updated 23 October 2026: Mac Studio M5 Ultra, DGX Spark, RTX PRO 6000 and DDR5 rows
See Also
- Can I run Mistral Large 4 locally? — the 1T MoE where 256GB holds only the 1-bit file
- Can I run Ling-3.1-flash locally? — the 560B MoE that also needs 200GB+
- Can I run MiMo V2.6 Flash locally? — another model that needs the 256GB tier
- Best local LLM for a Mac Studio M5 Ultra 256GB — what else runs on the box that runs Beam
- Which Mac Studio to buy for local LLMs — every M5 configuration priced
- Best models for the DGX Spark — 1 to 4-Spark setups
- Four RTX PRO 6000 Max-Q cards in one workstation — the 384GB route for official NVFP4
- —n-cpu-moe explained — how to put experts in system RAM
- Can I run Ornith 1.5 122B locally? — the model a 128GB box should run instead