Apache-2.0llmRuns locally501.3B
Reflection Beam: Specs, License & How to Run
Reflection 501B/23B Apache-2.0 sparse MoE for coding and long-horizon agents; weights opened 2026-10-16.
- License
- Apache-2.0
- Parameters
- 501.3B (active 23.4B)
- HF likes
- 148 (2026-10-19)
- First seen
- 2026-10-07
TL;DR. Reflection Beam is a llm model from Reflection. Parameters: 501.3B (active 23.4B).License: Apache-2.0. It is listed under ChatGPT, Closed LLM APIs, Cursor.
| Developer | Reflection |
|---|---|
| Type | llm |
| Architecture | Sparse MoE transformer; 52 layers (3 dense + 49 MoE); 256 experts, 8 routed per token; Gated GQA + Sliding Window Attention; Modified Harmony chat template |
| Parameters | 501.3B · active 23.4B |
| Modalities | text |
| Languages | en |
| Context | 1048576 |
| License | Apache-2.0 — Apache License 2.0 on the Hugging Face LICENSE file and card (license apache-2.0) as of 2026-10-19. Not legal advice. |
| Hardware | ~720 GB (4× B200 serving NVFP4, Hugging Face blog) — Hugging Face’s 16 October 2026 Beam blog says the BF16 checkpoint is about 1 TB in safetensors, NVFP4 shrinks that to ~280 GB of weights, and a practical NVFP4 serve is four NVIDIA B200s (720 GB) with room for long context. Two B200s (360 GB) can hold the weights with little room left. FP8 is named for Hopper/Ampere; vLLM BF16 is shown on 16 GPUs (two 8× H200 nodes). developers.reflection.ai serving-weights (checked 2026-10-19) says the NVFP4 model is 273 GiB. HF API usedStorage on the BF16 card is 1,002,668,503,017 bytes (~934 GiB on disk). forkalt has not measured peak VRAM. (not independently tested) |
| Hugging Face likes | 148 (as of 2026-10-19) |
| Model card | https://huggingface.co/reflection/Beam-1.0-501B-A23B |
| First seen | 2026-10-07 |
Reflection Beam (card name Beam-1.0-501B-A23B) is an ungated open-weight sparse Mixture-of-Experts language model from Reflection. Official hosts we opened on 2026-10-19: reflection/Beam-1.0-501B-A23B (BF16), NVFP4, FP8, DFlash2 drafter, the Apache 2.0 LICENSE, the Hugging Face post Welcome Beam (16 October 2026), Reflection’s Beam is now available (16 October 2026), the earlier Introducing Beam preview (5 October 2026), the model card, serving-weights docs, and Mirror CLI at reflection-oss/mirror.
This identification page hangs off ChatGPT alternatives, closed LLM APIs, and Cursor alternatives. Those mappings are editorial. We have no sourced quality comparison against ChatGPT or Cursor.
Head-to-head facts versus Meta’s gated US open-weight MoE: Beam vs Muse Spark Open.
What we verified (2026-10-19)
| Field | Value | Source |
|---|---|---|
| HF likes / downloads | 148 / 1,591 | HF API reflection/Beam-1.0-501B-A23B |
| Card created / last modified | 2026-10-07T19:52:01Z / 2026-10-16T13:22:56Z | same |
| Gated | false | HF API |
| License (card + file) | apache-2.0 / Apache License 2.0 | HF cardData + LICENSE |
| Pipeline tag | text-generation | HF API |
| Params (card) | 501.3B total / 23.4B active | Card “Model Overview” |
| safetensors.parameters | 501,267,612,352 | HF API |
| Context | 1M tokens | Card |
| Experts | 256; 8 routed per token | Card + Hugging Face blog |
| Layers | 52: 3 dense + 49 MoE | Card |
| Modalities | Text | Card |
| Knowledge cutoff | June 30, 2026 | Card |
| BF16 usedStorage | 1,002,668,503,017 bytes (~934 GiB) | HF API |
| NVFP4 sibling likes / downloads | 33 / 1,596 | HF API …-NVFP4 |
| FP8 sibling likes / downloads | 29 / 852 | HF API …-FP8 |
| DFlash2 likes / downloads | 30 / 432 | HF API …-DFlash2 |
| Mirror CLI stars | 180 | GitHub API reflection-oss/mirror |
| Weights public post | 16 October 2026 | Reflection blog + HF blog |
The 5 October 2026 “Introducing Beam” post described a waitlist and said weights would ship later that month under Apache 2.0. The 16 October 2026 posts are the weight-drop we treat as the public checkpoint.
Hardware (vendor / datacenter-class)
runsLocally: true here means the weights are an ungated download, not that a laptop GPU will load them.
Copied from the Hugging Face Beam blog and Reflection serving docs (checked 2026-10-19):
| Checkpoint | Weight size (vendor) | Serving note (vendor) |
|---|---|---|
| BF16 | ~1 TB safetensors (blog); ~934 GiB usedStorage (HF API) |
vLLM example: 16 GPUs / two 8× H200 nodes (blog) |
| NVFP4 | ~280 GB (blog); 273 GiB (serving-weights) | Practical: 4× B200 / 720 GB (blog). Two B200s hold weights with little room left |
| FP8 | not independently measured | Named for Hopper / Ampere; 8× H200 example in the blog |
Runtimes the HF blog names for day-0: transformers, SGLang, vLLM. Speculative decoding ships as DFlash2. forkalt has not loaded these checkpoints.
How to run (from official docs, not yet tested)
Serving steps live on developers.reflection.ai/serving-weights (NVFP4 download + SGLang / vLLM / TokenSpeed). The HF blog shows vLLM OpenAI-compatible serve commands for BF16 and FP8. Recommended sampling on the card is not copied here because the README does not publish a short sampling table in the extract we opened; use the card.
Mirror CLI (coding harness, not the model): docs say uv tool install reflection-mirror, then mirror in a project directory, default model Beam-1.0-501B-A23B. Platforms copied from the quickstart: macOS 15+ on Apple Silicon; Linux glibc 2.28+ on x86-64 or ARM64; native Windows not supported. forkalt has not run these commands.
License
LICENSE is Apache License 2.0. The card tag is license:apache-2.0. Mirror CLI is also Apache-2.0 on GitHub. Not legal advice.
The card’s usage paragraph says Reflection does not prohibit distillation, using outputs to train other models, fine-tuning, PEFT, quantization, or merging. That sentence is upstream, not a forkalt legal opinion. Read the LICENSE and Reflection’s usage policy.
What we did not verify
Vendor rows on the card (SWE-Bench Verified 81.6%, Terminal Bench v2.1 80.1%, GPQA Diamond 90.6%, DeepSWE v1.1 43.3%, and the rest of the BF16/FP8/NVFP4 table at temp=0.7, top_p=0.99) stay self-reported. The “approximately 3x fewer inference-time reasoning tokens” line is a vendor claim. Arena.ai Code Arena WebDev (open-source filter, 18 October 2026 snapshot on the page we opened) lists beam-1.0-501b-a23b as a preliminary open-source row (score 1437, overall rank 80 on that snapshot). That is a third-party leaderboard, not a forkalt bake-off.
Related chat / coding weights on this site: Muse Spark Open (gated Meta 816B/42B multimodal), DeepSeek-V4.1-Flash (MIT 1017 checkpoint), Kimi K3. Compare Beam vs Muse Spark Open.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| SWE-Bench Verified (BF16, vendor) | 81.6% (self-reported) | https://huggingface.co/reflection/Beam-1.0-501B-A23B |
| Terminal Bench v2.1 (BF16, vendor) | 80.1% (self-reported) | https://huggingface.co/reflection/Beam-1.0-501B-A23B |
| GPQA Diamond (BF16, vendor) | 90.6% (self-reported) | https://huggingface.co/reflection/Beam-1.0-501B-A23B |
Deploy options
- local: Hugging Face weights (BF16)
- local: Hugging Face weights (NVFP4)
- local: Hugging Face weights (FP8)
- local: DFlash2 drafter
- local: Mirror CLI (reflection-oss/mirror)
Comparisons
FAQ
Are Beam weights public and what is the license?
Yes. huggingface.co/reflection/Beam-1.0-501B-A23B is ungated (created 2026-10-07; last modified 2026-10-16T13:22:56Z; 148 likes / 1,591 downloads in the 2026-10-19 API check; pipeline_tag text-generation; card license apache-2.0). The LICENSE file is Apache License 2.0. Reflection’s “Beam is now available” post is dated 16 October 2026. Not legal advice.
How big is Beam, and is it multimodal?
The official card (checked 2026-10-19) lists 501.3B total / 23.4B active parameters, 52 layers (3 dense + 49 MoE), 256 experts, 1M context, text in / text out, English predominantly. Hugging Face safetensors.parameters sum to 501,267,612,352 (~501.3B). The card’s “Supported Modalities” row is Text. It is not a vision or audio model.
What is Mirror CLI?
Mirror is Reflection’s coding harness. github.com/reflection-oss/mirror is Apache-2.0 (180 stars, created 2026-10-12, last push 2026-10-16 in the 2026-10-19 GitHub check). Developer docs (developers.reflection.ai/coding-agents-quickstart) install it with `uv tool install reflection-mirror` and default the model ID to Beam-1.0-501B-A23B. Mirror is a harness, not the weights. forkalt has not run the CLI.
Does Beam match ChatGPT or Muse Spark Open quality?
We do not claim that. Card and blog benchmark tables are vendor/self-reported. Arena.ai Code Arena WebDev (open-source filter, 2026-10-18 snapshot) lists beam-1.0-501b-a23b as a preliminary row — that is a third-party leaderboard, not a forkalt ranking. See Beam vs Muse Spark Open for sourced head-to-head facts only.
Sources & verification
- https://huggingface.co/reflection/Beam-1.0-501B-A23B
- https://huggingface.co/reflection/Beam-1.0-501B-A23B-NVFP4
- https://huggingface.co/reflection/Beam-1.0-501B-A23B-FP8
- https://huggingface.co/reflection/Beam-1.0-501B-A23B-DFlash2
- https://huggingface.co/reflection/Beam-1.0-501B-A23B/blob/main/LICENSE
- https://huggingface.co/blog/reflection-beam
- https://www.reflection.ai/blog/beam-is-available
- https://www.reflection.ai/blog/introducing-beam
- https://reflection.ai/beam-model-card
- https://developers.reflection.ai/serving-weights
- https://developers.reflection.ai/coding-agents-quickstart
- https://github.com/reflection-oss/mirror
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.