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

Reflection released Beam's weights on October 16, 2026, eleven days after introducing the model. Its launch post calls it "our first open-weight model", built for coding, reasoning and long-horizon agent work. The weights are a 501-billion-parameter mixture-of-experts transformer with about 23 billion parameters active per token, a one-million-token context, and text in with text out. They ship under Apache 2.0, the permissive open-source licence, which means no revenue threshold, no attribution line in your product and no separate agreement for anyone.

The timing makes the licence the story. Meta released Muse Spark Open the day before, a larger model that takes images, video and audio, under a custom licence with an attribution rule and a revenue clause for model-as-a-service providers. We covered it in our Muse Spark Open release post. Two US labs shipped frontier-scale open weights in two days, and they made opposite licensing choices. The side-by-side comparison is in Reflection Beam vs Muse Spark Open; this post is about Beam on its own terms.

Every source below was read on October 19. Reflection's figures are Reflection's: the benchmarks were run by Reflection, at its chosen settings, and nobody outside the company has reproduced them. Artificial Analysis published its own independent index score for Beam on October 16; the companion comparison sets it beside Muse Spark Open's.

Reflection shipped four things together. The weights, on Hugging Face, in three numeric formats: BF16, the full-precision release; FP8, an 8-bit build; and NVFP4, NVIDIA's 4-bit format for its Blackwell GPUs. A small draft model called DFlash2 for speculative decoding, a technique where a cheap model guesses the next several tokens and the large model checks them in one pass. A model card with Reflection's final benchmarks. And a safety technical report setting out how the model was tested before release.

Around the weights sits what the launch post calls "an open operating system for agents": more than 40 partners and integrations, reference agent integrations, deployment containers and customisation pipelines. Reflection's own line is that "the model is the kernel". Day-one support covers Hugging Face Transformers, vLLM and SGLang, and the model ships with Mirror CLI, Reflection's open-source coding-agent harness, which pairs a Python front end with a Rust agent core.

Why the bundle matters: open weights alone leave the hard parts to the buyer. A team that downloads 273 GiB still has to pick an inference engine, tune it for the hardware and wire the model into a coding agent. Shipping containers, a drafter and a harness on day one removes weeks from that path. The failure case is assuming the bundle is finished. It is a first release from a new lab; treat the harness and partner integrations as starting points to test, not as production-hardened infrastructure.

A mixture-of-experts model splits most of its layers into many smaller networks, called experts, and a router picks a few of them for each token. Beam has 52 layers, 49 of them with 256 experts each, and the router selects 8 per token. Reflection's model card puts the totals at 501.3 billion parameters, 23.4 billion of them active. The card also lists a vocabulary of about 152,000 tokens, a modified Harmony chat format, and a knowledge cutoff of June 30, 2026.

The two numbers answer two different questions. Active parameters set the compute per token, and so the speed and the serving cost per request. Total parameters set the memory, because every expert has to be loaded even though only a few run at once. Hugging Face's write-up puts it plainly: "deployment needs the memory of a 501B model but runs at the speed of a 23B one."

The long context has its own memory trick. Most of Beam's layers use sliding-window attention, which only looks back over the previous 512 tokens; one layer in six, plus the last, attends to the whole sequence. Hugging Face calculates that this keeps the attention cache at about 4.6 GiB for a 128,000-token conversation, against about 26 GiB if every layer kept every token. The practical effect is that long agent sessions and concurrent users fit on the same GPUs that hold the weights.

The failure case is reading 23B as a laptop model. It is not. The active count tells you how fast a token comes out once the weights are loaded; it says nothing about whether they fit. Budget memory against 501B, and speed against 23B.

Apache 2.0 is a standard open-source licence that permits commercial use, modification and redistribution, with a notice file and a patent grant. It has no clause that changes at a revenue level, no rule that your product has to name the model, and no restriction on who may offer the model as a hosted service. Muse Spark Open, Kimi K3 and GLM 5.3 all carry custom terms instead; compare them in our audit of 2026 open-weight licences.

Reflection's model card goes further than the licence text needs to. It says Reflection does not prohibit distillation, using the model's outputs to train other models, fine-tuning, parameter-efficient adaptation, quantisation, merging, defensive security research on systems you are authorised to test, or good-faith adversarial evaluation. The card adds that it is informational and does not modify the licence, and points to a usage policy worth reading before a launch.

The failure case is assuming permissive means unreviewed. The weights are yours to use, but the data you run through them, the outputs your product shows and the sector rules you work under are still yours to handle. Apache 2.0 removes the vendor-terms question, not the compliance one.

Reflection's serving documentation lists the download sizes exactly: 933.81 GiB for BF16, 468.72 GiB for FP8 and 273 GiB for NVFP4. Its serving guides for SGLang, vLLM and TokenSpeed use the NVFP4 build and assume four NVIDIA B300 SXM GPUs linked by NVLink. The B300 is NVIDIA's Blackwell Ultra part, so this is current-generation data-centre hardware, rented or owned.

Hugging Face's guide fills in the alternatives. NVFP4 fits on four B300s. Two B200s could technically hold the weights, but with little room left for the cache, so the guide recommends four. On the older Hopper generation it serves the FP8 build on eight H200s. The BF16 reference runs across sixteen H200s, two nodes of eight.

The example that makes this concrete: an agency that already rents a four-GPU Blackwell node for image or video work can load Beam on it in NVFP4 with room for long sessions. A team on Hopper needs twice the GPUs for the 8-bit build. The failure case is the one our open-weight self-hosting guide keeps returning to: buying or renting for the weights and forgetting the cache and the headroom for more than one user at a time.

Quantisation stores weights in fewer bits to cut memory, and it usually costs some accuracy. Reflection published every benchmark in all three formats, which is the useful part of its card for anyone self-hosting: you can see what moving from BF16 to the 4-bit build costs on Reflection's own runs before you commit to hardware. All runs used temperature 0.7 and top-p 0.99.

On these rows the 4-bit build stays within about two and a half points of BF16, and comes out ahead on SWE-Bench Pro v2-Hard, which is noise around the same level rather than a gain. That matters because NVFP4 is the build that fits on four GPUs, so it is the one most teams will run. The comparison of Beam against other models on the same benchmark names, including where it trails, is in the companion comparison.

The failure case is treating these as your numbers. They are Reflection's runs on public suites, and a suite's score says little about your repository. Run your own tasks at the build and effort level you plan to serve, and keep the result next to the configuration that produced it.

Beam is a reasoning model: it writes hidden working before its answer, and you choose how much. Hugging Face's write-up lists five values, low, medium, high, xhigh and max, with medium as the default. A preserve_thinking option keeps that reasoning in context across turns, and tool calls use a dedicated grammar so the model's calls parse reliably. The tokenizer carries a response template that splits reasoning, answer and tool calls into separate fields.

The catch is that defaults differ by door. OpenRouter's listing for reflection/beam-1.0-501b-a23b, captured October 19, marks reasoning as mandatory with a default effort of max, as does Reflection's own API reference. A team that tries the model through OpenRouter at its default and then self-hosts at medium is comparing two different configurations, and paying for two very different numbers of reasoning tokens.

The example: an agent that classifies support tickets needs a short answer thousands of times a day, and low effort is the right setting for it. A refactoring agent on a large codebase earns high or above. The failure case is leaving max on for the first and paying for reasoning nobody reads. Set effort in the request, record it beside every evaluation, and do not rely on any default.

Most teams will try Beam without downloading anything. Reflection runs a first-party API in public beta, free to start, and has not published a price for it. Hosted inference partners serve it too. The one price list we can cite is OpenRouter's, where four providers listed the model on October 19 at prices that differ by a factor of five on input.

A fivefold spread for the same weights reflects providers picking different quantisations, hardware and margins. It is also a trap: a price quoted as "Beam costs $X" without a provider and a format means nothing. The failure case is benchmarking on one provider and deploying on another that serves a different quantisation. Pin the provider in your router settings once you have chosen.

Beam is text-only, predominantly English, tuned for coding and agent work, and licensed with no strings. That points at a clear set of buyers and leaves out others. The routing below is ours, built from Reflection's documentation and the licence, not from a trial of our own.

If you are choosing a model for an agent rather than a licence, our October release tracker lists the month's other major releases, including Meta's larger multimodal release. Help testing a self-hosted model against your own workload is part of our AI transformation work.