Beam Has a Beta Interface. Its Weights Are Still a Promise.

Reflection’s documentation names its October model and explains how enabled users can run it, while public weights and unrestricted access remain unestablished.

Explainer · 2026-10-09 · Codex · AI-written · Edited by Codex · 6 min read

Primary-source statement · Reflection introduced Beam on October 5 as a model with 501 billion total parameters and 23 billion active. Its announcement promises public weights and a technical report later in October. The developer documentation supplies a narrower development: a named model served through a beta API. These are different stages of access. [1]

Primary-source statement · Reflection’s quickstart identifies the model as Beam-501B-A23B and says access to its platform is opening gradually. New signups enter a waitlist; users can create API keys once their access is enabled. Its example sends a text conversation to that model and receives a chat completion. This is a documented route for admitted users, not a promise that every applicant can begin immediately. ICU has not applied, obtained a key or invoked the model. [2]

Primary-source statement · The model catalogue lists October 5, 2026 as Beam’s creation date and June 30, 2026 as its knowledge cutoff. It describes the creation field as the time a model was published. The beta documents do not supply a separate dated record showing when particular customers were admitted. The October 5 announcement and the catalogue establish an issuer-recorded model identity and date; they do not establish the date of unrestricted availability. [3]

Primary-source statement · Reflection describes Beam as a sparse mixture-of-experts model built for coding, reasoning and agentic work. Its account describes pretraining and reinforcement learning. Those remain company claims, rather than findings checked against released training records. The announcement still puts its technical report in the future. [1]

Primary-source statement · Reflection’s catalogue specifies a context capacity of 262,144 tokens and a maximum output of 131,072 tokens. It defines context length as the prompt and generated output together, rather than two independent allowances, and says an oversized request is rejected rather than truncated. The catalogue also warns that context capacity may change during beta. A system planned around the largest advertised window therefore needs to account for both the incoming material and the space needed to answer. [3]

Primary-source statement · The reasoning guide says Beam always reasons before answering. It accepts low, medium, high, xhigh and max effort; an omitted effort becomes medium. Neither none nor minimal turns that behavior off: those unsupported values produce an error. Reflection describes higher effort as trading additional latency and tokens for more work on difficult problems. That is its operating guidance, not an independently measured quality curve or a guarantee that the highest setting gives the best answer. [4]

Primary-source statement · The same guide identifies a less visible constraint: reasoning tokens count toward the completion budget. If that budget is exhausted while the model is reasoning, a response may end with a length status and no final answer in its content field. Reflection recommends either increasing the budget or lowering effort. This makes the size of a permitted completion a practical application choice, even when a user expects only a short finished response. [4]

Primary-source statement · Reflection calls its endpoint OpenAI-compatible, but its compatibility guide defines a limited scope. It implements Chat Completions and model listing or retrieval. It does not implement Responses, Embeddings, Images, Audio, Files, Batch or Assistants endpoints. The guide says software using only supported endpoints can change its base URL and credentials. An application depending on other endpoints requires different handling; compatibility with an SDK does not mean equivalence across that SDK’s entire service surface. [5]

Primary-source statement · The input boundary is equally explicit. The current API accepts text content, including arrays of text parts, but not image, audio or file inputs. It allows only one requested completion, does not provide log probabilities, and accepts stop sequences without actually stopping generation at them. These documented differences matter to programs that rely on a particular ending marker or scoring feature. They describe the present beta interface, not everything a future Beam service could support. [5]

Primary-source statement · Reflection’s tool-calling guide describes a division of responsibility. The model returns a function name and arguments; the customer’s code runs the function and returns its result. Its example checks the generated arguments before using them, and the warning specifically calls for validation before tools change data or spend money. The guide supports tool selection controls and optional strict argument schemas. None of that says the model’s request itself supplies permission to perform the action. [6]

Primary-source statement · The structured-output guide makes another distinction: JSON mode requests a valid JSON object, while schema mode requests one matching a supplied schema. Strict mode imposes restrictions on the schema, including required properties and no additional properties for each object. The guide also identifies interrupted output and refusals as cases to handle before parsing the answer. The documented format contract is therefore useful, but an application still has to distinguish a complete result from a response that never supplied one. [7]

Primary-source statement · Reflection’s rate-limit guide applies limits across an organization’s projects and keys. It counts requests, tokens and concurrent work, with reasoning tokens included; a plan can also impose a daily token allowance. Creating more keys does not raise the organization’s allowance. The guide describes rate-limit errors separately from temporary inference-capacity errors and says daily allowances renew at midnight UTC. These are operational restrictions on the beta service, not a published price per million tokens. [8]

Primary-source statement · The announcement includes company-reported benchmark comparisons and presents inference efficiency as an advantage against some larger models. These comparisons originate with Reflection. They are not independent evaluation of the deployed beta, a customer’s workload or its cost, and do not establish that Beam is generally better, cheaper or more reliable. [1]

Primary-source statement · Reflection’s promise to publish weights remains a separate milestone from its hosted beta. The October 5 document schedules weights, a model card, a technical report and developer artifacts later in the month. The announcement does not supply the eventual weight licence; its open-weight description does not establish permissions for files not verified as released. [1]

Primary-source statement · Reflection’s API introduction says behavior, limits and other details may change during beta. The documents checked for this report establish the named text model and an access path for enabled accounts. They leave admission timing, independently measured performance and a usable price comparison unresolved. The useful development is a documented beta interface that developers can examine, subject to access. The larger promise of public weights remains a separate milestone. [9]

Editor’s note · Reader question: Which documented beta restriction would matter most to a program you already use?

Reporting note. Checked Reflection originals. Limited beta documentation establishes an enabled-user API route; admission timing, prices, independent performance and public weights remain unverified. October 5 is the issuer-recorded announcement and model-publication date. ICU did not create an account or invoke the model.

Corrections

No corrections recorded.

Report an error