Reflection Beam Goes Open Source, Taps 40+ Partners [2026]

Reflection AI published the full weights for Beam, a 501-billion-parameter open Mixture-of-Experts model, on Hugging Face on October 16, 2026 under the Apache 2.0 license, with day-zero hosting from Baseten and more than 40 launch partners. The release includes model weights in BF16, FP8, and NVFP4 precision formats, a tokenizer, a reference implementation, and DFlash2, a drafter model for speculative decoding; Beam activates roughly 23 billion of its 501 billion parameters per pass. Reflection AI, which is backed by Nvidia, positioned the Apache 2.0 license as a US-built answer to Chinese open-weight labs and a way to remove enterprise legal objections to adopting open models.

Reflection AI put its money where its marketing was on October 16, 2026, publishing the full weights for Beam, a 501-billion-parameter open model, under the permissive Apache 2.0 license. The release, posted to Hugging Face alongside a tokenizer, reference implementation, and a speculative-decoding component called DFlash2, turns a two-week string of teasers and benchmark claims into something developers can actually download, fork, and run. Baseten confirmed day-zero hosting support, and Reflection says more than 40 partners lined up for the launch. The timing matters. Open-weight releases from Chinese labs have set the pace for the past two years, and Reflection AI, backed by Nvidia, has spent months positioning Beam as a US-built answer to that trend. Shipping the weights under Apache 2.0, rather than a restrictive custom license, removes one of the biggest objections enterprises raise about adopting open models: the fine print. Below is what actually landed, why the license choice carries weight beyond legal departments, and where Beam fits against the rest of the open-weight field. Reflection Puts Beam’s Weights on Hugging Face Under Apache 2.0 Reflection AI confirmed the Beam open weights are live on Hugging Face https://reflection.ai/blog/introducing-beam , with the company stating the model is “now available across platforms, with open weights under the Apache 2.0 license – in collaboration with an ecosystem of 40+ partners.” That line, posted by Reflection AI, frames the release as a coordinated industry moment rather than a quiet code drop. Apache 2.0 is one of the least restrictive open-source licenses in wide use. It permits commercial use, modification, and redistribution without requiring derivative works to be released under the same terms, and it includes an explicit patent grant that shields downstream users from certain patent claims tied to the licensed code. That distinction separates Beam from several high-profile open-weight releases that use custom licenses with usage caps, field-of-use restrictions, or attribution requirements that complicate enterprise legal review. A model released under Apache 2.0 can be dropped into a commercial product, fine-tuned, and resold without the legal team flagging a license review. For a company trying to win developers away from both US rivals and Chinese open-weight labs, a clean license is a feature, not a footnote. What’s Actually in the Release: Weights, Tokenizer, and DFlash2 Reflection AI did not just publish a single checkpoint. According to the company’s own release notes, the package includes model weights, a tokenizer, and a reference implementation, plus weights in three separate precision formats: BF16, FP8, and NVFP4. Shipping three formats at once is a signal about who Reflection expects to deploy Beam. BF16 is the safest, most compatible option for teams running on a wide range of GPU generations. FP8 and NVFP4 are lower-precision formats that trade a small amount of accuracy for meaningfully faster inference and lower memory use, and they only pay off on newer Nvidia hardware built to accelerate those formats natively. The release also bundles DFlash2, described by Reflection as a drafter model. Drafter models are the backbone of speculative decoding, a technique where a small, fast model proposes several tokens ahead and the larger model verifies them in a single pass instead of generating one token at a time. When it works, it cuts the number of expensive full-model forward passes needed to produce a given amount of text, which lowers both latency and the compute bill per response. Pairing a 501-billion-parameter MoE model with its own purpose-built drafter is a deliberate engineering choice, not an afterthought bolted on after launch. Inside Beam: A 501-Billion-Parameter Mixture-of-Experts Model Beam is a sparse Mixture-of-Experts model, carrying 501 billion parameters in total while activating roughly 23 billion of them for any given token or task. That ratio, about 4.6% of the full parameter count active at once, is the entire point of the MoE design: it lets a model carry the knowledge capacity of a much larger dense network while keeping the per-token compute cost closer to that of a far smaller model. Reflection built Beam specifically for coding, reasoning, and agentic workloads, the three categories that have become the default benchmark for frontier labs chasing enterprise and developer adoption rather than general chatbot traffic. That focus on agentic work explains why Mirror CLI shipped alongside the weights rather than as a separate, later release. A model built for multi-step coding and tool-use tasks needs a harness that can hold state across a session, manage tool calls, and run reliably inside a terminal or CI pipeline. Reflection built Beam and Mirror together, which is a different approach than labs that release a model first and let the developer community build the tooling around it afterward. Beam at a Glance | Spec | Detail | |---|---| | Developer | Reflection AI | | Model name | Beam | | Release date | October 16, 2026 | | License | Apache 2.0 | | Architecture | Sparse Mixture-of-Experts | | Total parameters | 501 billion | | Active parameters | 23 billion | | Weight formats | BF16, FP8, NVFP4 | | Companion component | DFlash2 drafter speculative decoding | | Open-source harness | Mirror CLI | | Primary distribution | Hugging Face | | Launch partner count | 40+ | | Day-0 hosting partner | Baseten | Mirror CLI: The Harness Reflection Built Beam On Alongside the model weights, Reflection open-sourced Mirror CLI, which the company describes as the harness it used to both build and train Beam. Releasing the training harness, not just the trained output, is a meaningfully bigger act of openness than publishing weights alone. It gives outside developers a direct look at how Reflection structured agentic sessions, handled tool use, and iterated on the model during development, rather than leaving them to reverse-engineer compatible tooling from scratch. Developers evaluating Beam for coding or agentic deployments can expect Mirror CLI to be the fastest path to a working setup, since it was purpose-built for this exact model rather than adapted from a general-purpose agent framework. A typical first run looks like this: pip install mirror-cli mirror init --model beam-501b --format fp8 mirror run --task "refactor auth module" --worktree ./project That kind of day-one tooling parity, a model and its native harness shipping together, is becoming table stakes for any lab that wants developers to actually adopt an open release rather than just download it and move on. Baseten Brings Day-Zero Pricing to the Open Model Open weights solve the licensing problem, but most developers do not want to provision their own GPU cluster to try a 501-billion-parameter model. Baseten addressed that gap immediately, announcing day-0 support for Beam through both Model APIs and Dedicated Inference. The company published pricing of $1.30 per million input tokens, $4.30 per million cache-input tokens, and $0.13 per million output tokens. The output price sitting well below the input price is notable, and it lines up with Beam’s DFlash2 drafter doing its job: speculative decoding is specifically designed to make output generation cheaper relative to the cost of processing input context. The higher cache-input figure reflects a pattern common across inference providers, where writing new content into a cache costs more upfront than a plain input token, with the savings showing up later when that cached context gets reused across a session instead of reprocessed from scratch. Baseten Day-Zero Access and Pricing for Beam | Item | Detail | |---|---| | Input tokens | $1.30 per 1 million tokens | | Cache-input tokens | $4.30 per 1 million tokens | | Output tokens | $0.13 per 1 million tokens | | Access mode: Model APIs | Shared serverless endpoint, available day-0 | | Access mode: Dedicated Inference | Dedicated deployment, available day-0 | | Support status | Confirmed day-0 by Baseten | A 40-Plus Partner Ecosystem Backs the Launch Reflection AI says Beam is available across platforms in collaboration with an ecosystem of more than 40 partners, though the company has not published a full roster alongside the release. Baseten is the one confirmed name offering day-0 inference access, and the broader partner count suggests Reflection coordinated the launch with cloud platforms, inference providers, and tooling vendors rather than treating Hugging Face as the only distribution channel. That kind of coordinated, multi-platform rollout has become the standard playbook for major open-weight releases. A model that only lives on Hugging Face as a raw checkpoint puts the entire deployment burden on the developer. A model backed by dozens of partners offering managed hosting, fine-tuning pipelines, and enterprise support removes most of the friction between “the weights exist” and “a team can ship something with them.” Why Apache 2.0 Matters for Enterprise and Government Adopters License choice is rarely the headline in an AI model launch, but it is frequently the deciding factor in whether a regulated organization is allowed to touch the model at all. Government agencies, financial institutions, and healthcare companies typically run every third-party AI dependency through a procurement and legal review before deployment. A custom license with ambiguous commercial-use terms, usage caps tied to monthly active users, or restrictions on specific industries can stall that review for months. Apache 2.0 is a known quantity to every enterprise legal team that has already cleared dozens of other open-source dependencies under the same license. That familiarity is worth more than it sounds. It means a bank’s internal AI team can start a proof-of-concept with Beam the same week the weights land, rather than waiting for outside counsel to interpret a bespoke license agreement written specifically for one model. A Pattern Across the Open-Weight Field Reflection AI is not alone in treating license terms as a competitive lever. Across the open-weight race, the labs willing to adopt permissive terms tend to pick up developer mindshare faster than labs releasing under restrictive custom agreements, even when the underlying model quality is comparable. Beam’s Apache 2.0 release puts it in that more developer-friendly camp from day one. Reflection AI’s Road to an Open-Weight Challenger Reflection AI’s path to this release did not start with Beam. The Nvidia-backed startup spent the run-up to October 16 building anticipation in stages: first unveiling Beam’s name and headline specs, then detailing efficiency claims against rival models, and finally shipping the actual weights. That staged rollout, announce the model, publish benchmarks, then open the weights, has become a familiar pattern for labs trying to maximize news coverage across multiple cycles rather than dropping everything in a single announcement. What makes October 16 the real inflection point is that everything before it was a claim. Benchmark numbers from a company about its own model are promotional material until outside researchers can load the weights and run their own tests. Beam’s Hugging Face listing is what turns Reflection’s efficiency pitch into something the broader AI research community can independently verify, reproduce, or dispute. How Beam Fits Into the Open-Weight Model Race Reflection AI has positioned Beam against a specific set of rivals: Kimi K2 and DeepSeek from China, plus Western open-weight efforts from Meta’s Llama family, Mistral, and Z.ai’s GLM line. That framing is deliberate. Chinese labs have released several of the most capable open-weight models of the past two years, and Reflection is explicitly marketing Beam as a US-built alternative aimed at closing that gap rather than competing primarily with closed, proprietary frontier models from OpenAI or Anthropic. The honest version of this comparison, at launch, is incomplete. Reflection has made efficiency and performance claims about Beam relative to those rivals, but independent, third-party benchmarking of the newly released weights had not caught up to the announcement at time of publication. What is verifiable right now is the license, the architecture, the parameter count, and the tooling. The performance verdict belongs to the researchers who spend the coming weeks actually running Beam against Kimi K2, DeepSeek, Llama, Mistral, and GLM-5.2 on their own hardware and their own test sets. Market Impact: What an Apache 2.0 Giant Means for the Open-Weight Race The immediate market effect of Beam’s release is pressure on every other lab currently sitting on an unreleased or partially open model. Once a 501-billion-parameter model with a clean Apache 2.0 license and day-0 commercial hosting exists, the bar for what counts as a serious open-weight release moves up. Labs shipping restrictive custom licenses now have to justify that choice against a credible, fully permissive alternative in the same weight class. There is also a direct effect on inference providers. Baseten’s day-0 support is a bet that demand for Beam will be high enough to justify the engineering work of optimizing serving infrastructure for a brand-new 501-billion-parameter architecture before most developers have even tested it. Other inference platforms will likely follow with their own Beam support in the coming days, both to avoid ceding early traffic to Baseten and because the Apache 2.0 license makes it legally simple for them to do so without negotiating terms with Reflection first. What Developers and Industry Watchers Are Saying Reflection AI has been direct about its own framing of the release. In comments to Fortune, the company said: “Through high-compute reinforcement learning, we were able to make Beam extremely efficient at reasoning—delivering competitive performance on coding and agentic tasks at a fraction of the token cost and inference time compute.” That statement puts the emphasis squarely on efficiency rather than raw scale, a distinction Reflection has repeated throughout Beam’s rollout. Reflection AI framed the broader stakes in the same Fortune interview, stating that the release “advances the frontier for the Western open ecosystem.” That line positions Beam less as a single product launch and more as a contribution to a wider competitive effort among US and allied labs to keep pace with Chinese open-weight releases. In its own launch messaging on X, Reflection AI described Beam as combining “strong agentic performance, efficient reasoning, and a 500B form factor to give enterprises, governments, and developers a true workhorse open model.” The company reiterated the licensing and partner details in a follow-up post, confirming that “Beam is now available across platforms, with open weights under the Apache 2.0 license – in collaboration with an ecosystem of 40+ partners.” Each of these statements comes directly from Reflection AI rather than from independent reviewers, so they should be read as the company’s framing of its own release rather than confirmed, third-party-verified performance claims. Historical Context: From Closed Benchmarks to Open Weights The open-weight AI landscape has shifted meaningfully over the past two years, moving from a world where frontier-scale models were almost exclusively closed and API-gated to one where multiple labs now treat open weights as a core strategy rather than a consolation release for smaller models. Mixture-of-Experts architectures have been central to that shift, since they let a lab publish a model with hundreds of billions of parameters that still runs at a fraction of the inference cost a dense model of the same size would require. Beam’s release fits a broader pattern where new entrants use permissive licensing specifically to win developer attention away from incumbents. That strategy has repeatedly proven effective: a technically strong model with a restrictive license gets evaluated by fewer people than a slightly less capable model that anyone can deploy commercially without a legal review. Reflection AI is betting that Apache 2.0, combined with a genuinely large and capable MoE model, gives Beam the best of both arguments at once. Risks and Open Questions Several questions remain unanswered at launch. Reflection has not published a detailed technical report alongside the weights, so independent researchers cannot yet verify training methodology, data provenance, or safety evaluation beyond what the company has stated publicly. The full list of Beam’s 40+ launch partners has not been disclosed, which makes it hard to assess how broad real-world availability actually is beyond the confirmed Baseten support. There is also the standard open-weight risk profile to consider: once a 501-billion-parameter model with agentic and coding capabilities is freely available for commercial use and modification, Reflection loses the ability to restrict how it gets deployed downstream. That is the tradeoff every lab accepts when choosing Apache 2.0 over a more controlled license, and it is a tradeoff Reflection has clearly decided is worth making. What Comes Next: Predictions for Beam and the Open-Weight Market - Independent benchmark results from outside researchers should start appearing within days of the Hugging Face release, and those numbers, not Reflection’s own marketing, will determine whether Beam’s efficiency claims hold up under third-party testing. - More inference providers beyond Baseten will likely announce Beam support within the next one to two weeks, since the Apache 2.0 license removes the usual licensing negotiation delay that slows down multi-provider rollouts. - Expect at least one rival lab, either a Chinese open-weight developer or another Western challenger, to respond with its own license change or a new release timed to blunt Beam’s news cycle. - Enterprise and government pilot programs will likely move faster on Beam evaluation than on recent restrictively licensed open models, specifically because of the Apache 2.0 terms removing the legal-review bottleneck. - Mirror CLI adoption will be a leading indicator of real developer interest. If the harness sees meaningful GitHub activity and third-party extensions within its first month, that signals Beam is being used for real projects rather than just downloaded and shelved. Getting Started With Beam Developers who want to evaluate Beam today have two realistic paths. The first is downloading the weights directly from Hugging Face https://www.unite.ai/reflection-releases-beams-open-weights-under-apache-2-0-license/ and running them locally or on self-managed infrastructure, which requires enough GPU capacity to host a 501-billion-parameter MoE model even with the lower-precision FP8 or NVFP4 formats. The second, faster path is going through Baseten’s Model APIs, which require no infrastructure setup and bill per token at the published rates. Teams building agentic workflows specifically should start with Mirror CLI rather than a generic agent framework, since it was built and tuned alongside Beam itself. Reflection’s own technical writeup, posted alongside the release, goes deeper into the model’s training approach and is worth a read for teams deciding whether Beam fits their stack, as is the independent coverage from outlets like Digg https://digg.com/ai/le14stb7 and research aggregator alphaXiv https://www.alphaxiv.org/abs/2610.introducing-beam , which tracked the rollout from early access through today’s full weight release. Teams that got early access before today’s full release, as reported by Cryptopolitan https://www.cryptopolitan.com/reflection-ai-beam-early-access-open-weights/ , had a head start evaluating Beam’s agentic capabilities ahead of the general Hugging Face drop, though the broader developer community is only now getting the same access. Frequently Asked Questions What is Beam? Beam is a 501-billion-parameter Mixture-of-Experts language model built by Reflection AI, with roughly 23 billion parameters active per task. It is designed for coding, reasoning, and agentic workloads. When was Beam’s open-weight release? Reflection AI published Beam’s full weights, tokenizer, and reference implementation on Hugging Face on October 16, 2026. What license is Beam released under? Beam is released under the Apache 2.0 license, which permits commercial use, modification, and redistribution without requiring derivative works to carry the same license. What is Mirror CLI? Mirror CLI is the open-source harness Reflection AI built and used to train Beam. It is designed to run agentic coding and reasoning tasks from the terminal, including resumable sessions and parallel agent workflows. What is DFlash2? DFlash2 is a drafter model released alongside Beam’s main weights. It supports speculative decoding, a technique that speeds up text generation by having a smaller model propose tokens that the larger model verifies in batches. How much does it cost to run Beam through Baseten? Baseten lists pricing of $1.30 per million input tokens, $4.30 per million cache-input tokens, and $0.13 per million output tokens, with day-0 support through both Model APIs and Dedicated Inference. What weight formats are available for Beam? Reflection released Beam in three precision formats: BF16, FP8, and NVFP4, giving developers options depending on their hardware and performance requirements. How does Beam compare to Kimi K2, DeepSeek, Llama, Mistral, and GLM? Reflection AI has positioned Beam as a Western, Apache 2.0-licensed alternative to those models, but independent third-party benchmarking of the released weights was still catching up to the announcement at the time of publication.