The Reflection AI Beam model is now downloadable. Reflection AI, the Nvidia-backed startup behind it, posted the open weights for the 501-billion-parameter model on Hugging Face on October 16 under an Apache 2.0 license. The company is pitching Beam as a US-built alternative to the Chinese open models that many American companies already run, mostly because those models cost less.
The pitch targets a specific buyer. VentureBeat framed Reflection's argument as fewer tokens per task rather than cheaper tokens, and it noted that Chinese labs have held the top of open-model performance rankings. That distinction shifts the debate from price to efficiency, a claim that only independent testing can settle.
Which Beam weights are public and where to get them
Reflection introduced Beam on October 5 as its first open-weight model. The company's introduction describes a sparse mixture-of-experts design with 501 billion total parameters and 23 billion active per token, built for coding, reasoning and agent work. Reflection says the model was trained on 23.8 trillion pretraining tokens, and AI Weekly reports a 1-million-token context window.
Access started with an early-access program and a waitlist, with the weights promised later in the month, as TechCrunch reported at launch. On October 16, Reflection posted on X that "Beam is now available across platforms, with open weights under the Apache 2.0 license," in collaboration with more than 40 partners. According to Shattered, the release includes:
- A tokenizer and a reference implementation
- BF16, FP8 and NVFP4 weight variants
- DFlash2, a speculative-decoding drafter
- Mirror CLI, the harness Beam was trained against, which Reflection calls an "open operating system"
Day-zero support was announced for Hugging Face transformers, vLLM, SGLang, LM Studio and Unsloth. Baseten also confirmed day-zero hosting, which gives teams without their own GPU clusters a way to test the model the day it went public.
How Beam's benchmark claims compare with GLM-5.2
Reflection says Beam is on par with Z.ai's GLM-5.2 and close to Alibaba's Qwen3.8-Max, according to The Hill. The company also claims roughly 80% on Terminal-Bench 2.1 at about 25,000 tokens per attempt, compared with 80,000 tokens for GLM-5.2, as AI Weekly summarized. On one test, Reflection says Beam uses about a third of GLM-5.2's tokens.
Treat those figures as company claims. Reflection says reproducible evaluation recipes are "coming soon," and no independent verification has been published. Coverage is mixed as well. One early write-up says Beam trails Moonshot's Kimi K3 on agentic coding.
Reflection has not directly compared Beam with Muse Spark Open, the Meta model released a day before Beam's weights went live. Meta pitched Muse Spark Open as the most capable US open-weight model, according to TechCrunch. The table below lists only the figures reported so far for both models.
| Beam (Reflection AI) | Muse Spark Open (Meta) | |
|---|---|---|
| Launch | Introduced Oct. 5, weights posted Oct. 16 | Released Oct. 15 |
| Total parameters | 501B, with 23B active per token | 816B |
| Context window | 1M tokens | 1M tokens |
| License | Apache 2.0 | Not stated in the reporting reviewed |
| Company pitch | Coding, reasoning and agent work; fewer tokens per task | Most capable US open-weight model; designed to run on Apple silicon |
Why a US-built open model matters for buyers
For developers and enterprises, the most immediate change is legal. Apache 2.0 is a permissive license that legal teams already know how to review, which matters for sovereign and regulated buyers who have been wary of Chinese-origin models. The Hill reported that cheap Chinese open models have drawn national-security scrutiny, and that concern is the opening Reflection is aiming at.
The broader picture is a split in release strategy. OpenAI, Google and Anthropic keep their strongest models closed, while Chinese labs publish downloadable weights, according to Technology Org. Policy analysts in Washington reportedly say Chinese open models trail the US frontier by months. They also warn that heavy adoption creates technical dependence on Chinese model families.
Reflection was founded by two former Google DeepMind researchers and reportedly raised at a $25 billion valuation before launch. The company describes Beam as a "powerful workhorse model" for coding and agentic tasks, a phrase The Hill attributed to Reflection's blog. The dependence problem also reaches below the model layer. AI labs are trying to reduce reliance on a single chip supplier, a theme USA 24 covered in its report on Microsoft's Maia 300 chip.
The winners are easier to name than the losers. Compliance teams get a US-origin option under a permissive license. Chinese providers lose ground on price only if Beam matches them on quality per dollar, and that has not been shown yet.
What it takes to run a 501-billion-parameter model
Only 23 billion parameters activate for each token, which lowers compute per step. Every expert still has to sit in memory, though, so the design does not shrink the storage bill. Using the weight formats Reflection published, and ignoring the KV cache and runtime overhead, a rough estimate for the weights alone looks like this:
- BF16: about 1 terabyte (501 billion parameters at 2 bytes each)
- FP8: about 500 gigabytes
- NVFP4: about 250 gigabytes
That is multi-GPU territory for most teams. Small companies are more likely to reach Beam through hosted endpoints than through their own servers. Pricing from hosts beyond the day-zero partners has not been published, so a clean cost comparison is not yet possible.
How to test Beam before you switch
Teams evaluating Beam should treat the vendor numbers as a hypothesis to test. A practical sequence looks like this:
- Pull the weights from Hugging Face, or start with a hosted endpoint on Baseten to skip the hardware step. Choose the variant that fits your hardware, with NVFP4 the smallest footprint.
- Run Reflection's Mirror CLI harness first, since Beam was trained against it, so your early results match the setup the company describes.
- Build a private evaluation from your own tickets, code repositories or terminal tasks. Score pass rate alongside tokens per completed task, latency and cost, because Reflection's argument rests on token efficiency.
- Run the same tasks through GLM-5.2, Qwen or whatever model your team uses now, with identical prompts and tools.
- Send the Apache 2.0 text to counsel and check it against your procurement, security and export rules before any production rollout.
What to watch next
Reflection says it is already training a follow-on model, which signals a fast release cadence, according to AI Weekly. The near-term checkpoints are the reproducible evaluation recipes, independent benchmark results and pricing from hosts beyond the day-zero partners. Whether the Apache 2.0 license clears enterprise procurement and national-security reviews is still untested.
Policy is moving too. Massachusetts lawmakers are weighing independent safety tests for frontier models, and USA 24 has covered that proposal. Open-weight releases such as Beam and Muse Spark Open will test how any such rules apply when anyone can download a model.
The next real test is independent. If outside evaluators reproduce the efficiency claims, American developers will have a credible open option for coding and agent work. If they do not, Beam will still have put a US open model on the shortlist, and Reflection's next release will be judged on the same terms.
