Reflection released Beam's downloadable model weights on October 16, turning its earlier private preview into a choice between running the model on an organization's own infrastructure and calling a hosted API. The release announcement links the weights, a model card and a safety report. Its launch also names coding agents, serving systems and customization tools. For developers weighing those routes, the artifacts and access conditions matter more than the length of the partner list.

The big change

  • What changed: Beam's October 16 release made Apache 2.0 model weights available in three documented precision formats after an earlier private preview. Developers can inspect and host a specific checkpoint rather than rely solely on preview access.
  • Why it matters: Reflection's public-beta API offers another route through a key and a documented chat endpoint. Its current free allocation is capped daily, and the public terms leave the exact allowance to the account dashboard.
  • What to watch: Self-hosting, hosted inference and an agent harness require different components and budgets. The released drafter is optional and depends on Beam; the partner and training claims need route-specific verification. No independent Beam deployment or benchmark was conducted for this article.

What is available to download

Beam is a text-in, text-out instruction model. Its model card specifies a sparse mixture-of-experts architecture with 501.3 billion total parameters and 23.4 billion activated for a token, a one-million-token context window, and BF16, FP8 and NVFP4 numerics. Reflection's verified Hugging Face organization lists separate repositories for the BF16 model, FP8 model and NVFP4 model, each marked Apache 2.0. The card describes coding, reasoning and agent tasks as its intended uses.

The active-parameter count describes the part of the model used for a token. It does not describe the total weights a server must hold, its working memory, or the memory used by a long context. The one-million-token specification likewise does not establish that a particular deployment can serve that length at a useful speed or price. Teams choosing to self-host still need to select a checkpoint, compatible serving software and hardware capacity. This article did not download or run Beam, so it offers no measured memory or latency figure.

There is also a separate DFlash2 drafter repository. Its card says the drafter runs beside Beam in a speculative-decoding server; Beam verifies drafted tokens. It is not a standalone assistant. Reflection's reported decode-speed results for that pairing use Beam NVFP4 with tensor parallelism of four on B300 GPUs and specific workloads. They are vendor measurements, not a speed estimate for another machine or hosted service.

The API offers a different starting point

Reflection's API quickstart documents an account, an API key and an OpenAI-compatible chat-completions endpoint with model ID Beam-1.0-501B-A23B. The request takes text messages and returns a chat completion. That lets a developer evaluate the hosted model without arranging a self-hosted inference stack. Reflection calls the first-party API a public beta and its beta pricing terms say access is currently free with daily input and output allocations. The terms do not give a fixed allocation on the public page; the account dashboard shows the remaining balance and reset time. Requests are declined once an allocation is exhausted, and Reflection says it may change the limits or end the beta.

Those terms make an API key useful for an initial evaluation, but “free to start” is not a promise of unlimited free inference or a production price. A team planning a service would need its own account's limits and current commercial terms. The downloadable weights carry a separate hosting cost determined by the hardware and provider chosen; this reporting did not price a deployment.

A model release is not the whole agent stack

The October 16 announcement presents Beam alongside Reflection's Mirror CLI, other coding harnesses, serving integrations and training partners. Mirror has its own documented quickstart and is described as an early beta. Its coding-agent setup connects it to Beam through an API key; it also documents Pi, OpenCode and Hermes configurations. Mirror's quickstart lists macOS and Linux support and says native Windows is unsupported. These are agent applications and connection paths around the model, rather than capabilities that appear merely by downloading weights.

Reflection also names BF16, FP8 and NVFP4 serving support and a customization ecosystem. The separately published model repositories substantiate the three weight formats. The announcement's broader claims about partner availability and training workflows should be checked against each provider or recipe before committing to a build. In particular, its self-distillation recipe was described on October 16 as planned for release “in the coming days”; that launch wording alone does not establish a usable released recipe. This article treats the release as a set of inspectable model artifacts plus documented access routes, rather than certifying every named integration.

The model card's benchmark table reports, for example, 81.6% on SWE-bench Verified for BF16 and 80.2% for FP8 at temperature 0.7 and top-p 0.99. These are Reflection's published evaluations, with no BIG CHANGE rerun. A model card result does not determine how Beam will perform in a particular repository, agent harness or serving setup. Reflection has published a Safety Technical Report with its evaluations and release rationale. The Hugging Face cards separately label a general technical report “coming,” so the safety report should not be mistaken for that pending artifact.

Sources & further reading