Reflection AI Beam: A 501B Open-Weight Model Under Apache 2.0

Reflection AI Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active per token. Reflection published the weights on Hugging Face on 16 October 2026 under Apache 2.0, in BF16, FP8 and NVFP4. The benchmark figures come from Reflection, and the pages we read give no API price and no GPU requirement. This post separates what is stated from what is claimed.

Beam is worth a look mainly because the licence is Apache 2.0, which is the easiest kind for a legal team to clear. A permissive licence says what you may do. It says nothing about whether the model is good at your task.

We read Reflection's two posts and the Hugging Face pages on 9 November 2026. Where a number is ours, we say so.

What is Reflection AI Beam?

Beam is Reflection AI's first open-weight model. The company announced it on 5 October 2026, when it said Beam was still "undergoing final red-teaming and evaluations" and was available to "a select group of users". It released the weights on 16 October.

Here are the facts we read on the Hugging Face model card (repository reflection/Beam-1.0-501B-A23B) and in the two Reflection posts.

Reflection also gives training detail: 23.8T pretraining tokens, and reinforcement learning over more than 100 million rollouts with a maximum context length of 256K tokens. Those are Reflection's own figures.

The Hub's Beam 1.0 collection has four repositories: BF16, FP8, NVFP4 and a small drafter model, DFlash2, for speculative decoding. Reflection's release post names serving support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and Dynamo, and says Dell and HPE deliver Beam "behind the firewall" for on-premises and air-gapped sites. It lists AWS, Cloudflare, Google Cloud, Microsoft Foundry, OpenRouter, Together AI and others as launch hosts.

Reflection also runs a first-party API in public beta. The post says "It's free to start." It lists no price, and neither do we.

What has Reflection claimed, and what is vendor-reported?

Reflection's headline efficiency claim is that Beam scores "comparable to GLM-5.2" on advanced reasoning benchmarks while using 3 to 4 times less inference compute. The footnote matters. Reflection says the figures "represent an approximate compute comparison rather than measured inference cost."

So the 3 to 4 times figure is a vendor estimate of compute, not a measured bill. We cannot turn it into a cost per task, and we would not budget on it.

The benchmark table is also Reflection's. The announcement carries an update dated 27 October 2026 saying the tables reflect Beam's latest results. We copied a few rows below for the models most buyers will compare against. "NR" means the page did not report a score.

Benchmark (vendor-reported by Reflection, as updated 27 October 2026):

The page also gives Beam 80.9 on SWE Bench Verified, where it prints no figure for these three rivals.

Read the table as Reflection's own presentation. Beam is ahead of GLM 5.2 on three rows (SWE Bench Pro v1, MCP Atlas and AA-LCR) and behind on the other three shown. It trails GLM 5.3 and Qwen 3.8 Max wherever both report a score. That fits the word "comparable" in the efficiency claim, which compares with GLM 5.2, not with the newer GLM 5.3.

None of these runs is independent. We have not seen an outside party reproduce them. Our standing advice from the open-weight comparison applies: a benchmark run against a vendor's hosted endpoint is not the same artefact as a checkpoint you download and serve yourself.

Is Beam really open source?

Beam is open-weight, and its licence is an open-source one. Whether you call the model "open source" depends on a definition, and the definitions differ.

We read the LICENSE file in the Beam repository. It is titled "Apache License, Version 2.0, January 2004", and the Hub tags the repository apache-2.0. The licence text on apache.org contains no clause about hosting or serving software to third parties. Its grant covers the right to "reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute". It also includes a patent licence that terminates for a licensee who starts patent litigation over the work. We did not compare Reflection's file line by line with the Apache text, so treat the match as a reading, not a diff.

That is a different position from several other large open-weight releases. Mistral Large 4, Kimi K3.1 Pro and Muse Spark Open each add a condition aimed at businesses that sell model access to others. We read all three licences and compare them in our licence guide. Beam's file has no such condition.

The Open Source Initiative's Open Source AI Definition, version 1.0, asks for more than a licence on the weights. It says the preferred form for making modifications includes data information, the complete training and inference code, and the parameters. It adds that releases only count as open-source weights if they "must include the data information and code used to derive those parameters."

Reflection's announcement promised "weights, technical report, model card, and developer artifacts". We found no statement that the training data or the full training code is published. Under the OSI definition, that means we cannot call Beam open source AI. "Open-weight under Apache 2.0" is the accurate description.

For most buyers the practical question is narrower. Can we run it, modify it, fine-tune it and offer it inside our own product without asking anyone? On the licence we read, yes. A lawyer should confirm that for your use. This is not legal advice.

How much GPU memory does Beam need?

Reflection gives no GPU count and no memory figure. The model card has no hardware section, and the release post says only that the lower-precision formats "reduce memory requirements and accelerate token generation out of the box." What follows is our arithmetic.

Weights set the floor. Multiply total parameters by bytes per parameter. We use 501.3B and decimal gigabytes.

• BF16, 2 bytes per parameter: 501.3B x 2 = about 1,003 GB.
• FP8, 1 byte: 501.3B x 1 = about 501 GB.
• NVFP4, about 0.5 byte: The four-bit payload alone is 501.3B x 0.5 = about 251 GB. Block scales and higher-precision tensors add to that: the released NVFP4 repository shows 294 GB of files on the Hub. The Hub's parameter label is not a memory measurement, so size the service from its tensors, then add KV cache and runtime memory.

All 501B parameters must sit in memory, even though only 23B are active per token. Mixture-of-experts models save compute per token, not storage. We explain that distinction at more length in our GPU requirements post for Kimi K3.

Now compare with real hardware. NVIDIA's H200 page lists 141 GB of memory per GPU. NVIDIA's DGX B200 page lists 1,440 GB of GPU memory in total for eight GPUs. Headroom is node memory minus weights, which is what remains for the KV cache, activations and the serving runtime.

Two cautions go with the table. First, a node that fits the weights is not yet a usable service. Long contexts need KV-cache memory, and at 1M tokens that can become a large share of the total.

The KV-cache formula is: 2 (keys and values) x layers x KV heads x head dimension x bytes per value x tokens held. The Beam card gives the layer count (52) but not the KV head count or head dimension. It also uses sliding-window attention, which keeps fewer tokens per layer. So we cannot give a Beam figure, and we will not invent one. Ask Reflection or read the config file before you size a node.

Second, check format support. NVFP4 is a recent NVIDIA 4-bit format. We did not verify which GPU generations Beam's NVFP4 kernels support, so confirm that before planning around the smallest weights on older hardware.

Beam vs Qwen 3.8 vs GLM: which open-weight model fits?

The three families differ most on licence and size. We use only figures from model cards or earlier Swfte posts.

Here is how we would read it.

You need a small model you can run on modest hardware: Qwen3.8-27B. It is dense, Apache 2.0, and does not need a multi-GPU node. Our self-host guide for it covers the setup. Beam is the wrong tool here.

You need the cleanest licence at large scale: Beam, and the 27B Qwen at the smaller end. Both are Apache 2.0 on the files we read. The larger Qwen releases are custom licences. The Qwen Community License 1.0 asks anyone running a Model-as-a-Service or "AI Work Assistant" business for a separate licence, and the Qwen3.8-Max License adds a US$50 million revenue threshold to the same idea. We read both files.

You want the strongest benchmark scores: on Reflection's own table, Qwen 3.8 Max and GLM 5.3 score higher than Beam on most rows. Qwen3.8-Max is a hosted service. The downloadable Qwen3.8-2.4T-A95B checkpoint carries the custom Qwen3.8-Max licence, and hosted access follows the provider's service terms, which we did not read. That table is vendor-reported, so run your own tasks.

You care about active parameters and serving cost: Beam's 23B active is mid-sized for its total. Flash-Next activates 6B. GLM-5.3 activates 40B. Active parameters drive speed per token, and total parameters drive memory. Neither tells you quality per task.

We do not name a winner. Your task, hardware and legal review decide, and only a fixed test set can rank them.

What is still unknown about Beam?

Several things were not in the pages we read.

• No price for Reflection's own API. The release post says "free to start" and gives no rates. Hosts such as AWS and Together AI list Beam, and we did not check their prices.
• No hardware guidance. No GPU count, memory figure or throughput number from Reflection.
• No independent benchmarks. We have no third-party reproduction of the vendor table.
• KV-cache cost at long context. Not published, as covered above.
• Training data and code. No statement that either is released.
• Behaviour on your workload. Reflection's benchmarks centre on agentic coding, reasoning and tool use. The pages say little about extraction, summarisation or multilingual work, and the model is text only.

Reflection's post list also shows a safety post dated 5 November 2026, which we did not read. If you are in a regulated sector, read it before deployment.

Who should evaluate Beam now?

Evaluate it this month if you meet two conditions. You already run large open-weight models, or you have a legal reason to run inference on your own infrastructure. And your workload is agentic coding, tool calling or reasoning, which is where Reflection's numbers are aimed.

Wait for outside numbers if you run a hosted API today and would only switch for cost. The 3 to 4 times claim is a compute estimate. Until someone measures serving cost on common hardware, it is a reason to test and not a reason to migrate.

A sound evaluation runs in five steps:

1. Fix 50 to 100 real tasks from your own traffic, with a pass or fail check for each.
2. Run Beam through a hosted partner first. That tests quality without buying a node.
3. Run the same set on your current model, with the same prompts and sampling settings.
4. Measure tokens per task, because reasoning models can use many.
5. Only then price a self-hosted node, using the memory arithmetic above and Reflection's config.

If you plan to self-host, put the model ID behind an alias in your own code so a later swap is a config change. Our risk guide for regulated teams lists the controls to put around a downloaded model, and the GPU post for trillion-parameter models extends the memory method to larger releases.

Where Swfte sits

Beam is a model, and a model is one input to a stack. Swfte does not claim to host Beam, and we have not tested it.

Swfte Connect offers one API across many providers, with smart routing that optimises for cost and availability. You can bring your own keys from any provider, or use Swfte's pooled access. When a provider is out of funds, rate-limited or down, Connect can fall back to another provider in the chain. When a workspace reaches a usage cap, Connect can move its requests to a cheaper model.

That pattern suits an evaluation like the one above. You can keep your current model as the default, send a slice of traffic to a candidate, and keep a fallback in the chain. Models not in the catalogue are treated as bring-your-own-key with an optional base URL. A custom OpenAI-compatible endpoint, such as your own vLLM server running Beam, can therefore sit behind routing, failover, usage caps, cost tracking and the audit event stream.

For regulated buyers, the enterprise deployment generator produces Terraform, Helm and Docker Compose bundles. Generating one requires a licence key, and dedicated deployment is available on request.

The rule: test Beam on your own tasks through a host, and size hardware after the quality result is in.

---

Related: Open-weight licences compared, What a trillion-parameter open model needs, The best open-weight LLM in 2026, Open weights for regulated teams.

Sources: Reflection AI, Introducing Beam (5 October 2026, updated 27 October) · Reflection AI, Beam is available (16 October 2026) · Hugging Face, reflection/Beam-1.0-501B-A23B model card and LICENSE · Hugging Face, Beam 1.0 collection · Hugging Face, Beam NVFP4 repository files · Apache Software Foundation, Apache License 2.0 · Open Source Initiative, Open Source AI Definition 1.0 · NVIDIA, H200 · NVIDIA, DGX B200 · Hugging Face, Qwen3.8-27B · Hugging Face, Qwen3.8-Flash-Next · Hugging Face, Qwen3.8-2.4T-A95B · Hugging Face, Muse Spark Open card, accessed 9 November 2026.

Benchmark figures are Reflection's own and are not independently reproduced. Memory figures are our arithmetic from published parameter counts and NVIDIA's page figures, and exclude the KV cache and runtime overhead. Pages were read on 9 November 2026 and may change. This is not legal advice.