The promise, kept in the window it named

On October 5 this site's verdict was simple: Beam was an announcement, not a download. The weights were a date on a calendar, the model was still in red-teaming, and there was no public checkpoint anyone could check the claims against.

That objection expired on October 16. The Hugging Face repositories are public, the licence is the Apache 2.0 text Reflection promised, and the release includes the things a serious open release should include: a model card with final numbers, a safety technical report, three precision formats, a speculative-decoding drafter, and serving support in the frameworks people already use. The safety report's existence does not by itself establish thorough red-teaming or deployment safety; it is a document to read, not a clearance. In a month when another Western lab's open model is still a preview API with weights due later, shipping within the promised later-in-October window is itself information.

Read the final card, not the launch post

The final numbers are close to the preview numbers, but they are not the same numbers, and the differences cut both ways. SWE-bench Verified moved up to 81.6% for the BF16 build, from the 80.9% figure carried in preview coverage and still cited in Baseten's launch summary. DeepSWE v1.1 moved down, to 43.3%. Terminal-Bench v2.1 held at 80.1%.

The cheaper formats need a benchmark-by-benchmark reading, not a single discount. On Terminal-Bench v2.1, FP8 scores 78.9% and NVFP4 78.4%, which is 1.2 and 1.7 points below BF16, a Terminal-Bench-specific cost rather than a model-wide quantisation tax. On SWE-Bench Pro v1 (65.9% in FP8 and 65.6% in NVFP4, against 66.1% in BF16) and DeepSWE v1.1 (43.0% and 42.7%, against 43.3%) the cheaper builds also trail, and they trail on SWE-bench Verified too: 80.2% in FP8 and 80.7% in NVFP4, against 81.6% in BF16. But on SWE-Bench Pro v2-Hard the order inverts: FP8 scores 78.4% and NVFP4 78.6%, both above BF16's 76.5%. The headline BF16 result is not uniformly the highest, and which build leads depends on the task. NVFP4 will likely suit many hosted and capacity-constrained deployments, but suitability depends on hardware support, kernel and serving-stack support, latency and throughput needs, and these task-specific trade-offs, not on a universal ranking of the three formats.

One claim changed shape between the two documents, and changed units with it. In October the pitch was efficiency measured in compute: comparable performance to GLM-5.2 at roughly three to four times less inference compute. The final card frames it in tokens: roughly three times fewer reasoning tokens than GLM-5.2 for similar performance on selected long-horizon coding and agent tasks. Tokens and compute are related but not interchangeable, and reasoning tokens exclude prompt, tool, retrieved-context and other generated tokens, so the token figure is not a total inference-cost saving. A reader comparing Beam's efficiency against a rival's should check which unit each side is quoting.

All of these evaluations are still Reflection's. The card states its sampling settings, which helps. Launch partners including Artificial Analysis and Arena had pre-release access and ran launch testing, which is partner testing, not independent validation; independent post-release evaluation is still pending. That was true of every open model this year. It is not a reason to dismiss Beam, and it is a reason to treat 81.6% as a claim awaiting replication rather than a ranking.

One more limit is worth naming. The card describes the training data only in general terms, and the training pipeline is not released, so data provenance, contamination and reproducibility cannot be fully audited from this release. The self-distillation recipes that would help others reproduce the training approach were only promised for the coming days.

Downloadable is not the same as runnable

Apache 2.0 means anyone may download Beam. The arithmetic decides who can run it. At 501 billion parameters, the BF16 weights alone come to roughly 1,002 GB of storage; FP8 is about 501 GB, and Hugging Face lists the NVFP4 repository at a 255B size. Those are storage figures, not serving figures: neither raw file size nor parameter count gives a minimum GPU memory or a runnable configuration. Actual needs also include the runtime, activations, the KV cache, concurrency, the context length being served and the serving stack itself. The card lists a 1M-token context, but that alone does not prove effective long-context performance, usable throughput or affordable serving at a full window; retrieval and reasoning across a million tokens still need independent evaluation.

So the realistic audience splits in two. Enterprises and governments get what Reflection is actually selling: a permissively licensed model they can place behind a firewall through Dell or HPE, fine-tune on their own data, and, as the card states explicitly, distil or use to train other models without asking permission. Reflection is US-based, which is the qualified sense in which Beam is an American open model; the hosting, hardware, supply chain and deployment behind any given installation may not be wholly American. Everyone else gets a hosted endpoint, at which point Beam competes on price per token against the Chinese open models it was built to answer, and the only scores on the table are its maker's.

One detail in that hosted market deserves a second look before anyone budgets. Baseten's library page, accessed on October 16, lists Beam at $1.30 per million input tokens, $4.30 for cache input and $0.13 for output. Cache input priced above ordinary input, and output priced lowest of the three, inverts the pattern of every comparable price card. It may be a page or row error, or an unusual pricing structure. Treat it as unverified for budgeting until Baseten confirms it, and derive no value or cost comparisons from it.

The acquisition question, restated

This site reported the position on October 10, and it needs stating carefully. Nvidia was already a major investor in and backer of Reflection. Reuters, reporting on a Financial Times report, said talks were under way that could involve Nvidia investing further or acquiring the company. The deal status remains unconfirmed. Nvidia does not own Reflection, and this week's release changes nothing about that.

What the Apache 2.0 release does change is narrower and more durable than the takeover framing. Rights already granted under Apache 2.0 for copies lawfully obtained are irrevocable for those copies and cannot be retroactively withdrawn from them. That is not a guarantee of continued future distribution, updates, support, trademarks or access: an acquirer would control the company, its future releases, its repositories and its support. The protection covers the copies already out, not the pipeline behind them.

The weights shipped first, within the promised window, under the licence Reflection named. Public weights permit inspection and replication attempts, but replication is not a download away: it requires major compute, the correct quantised artefacts, a serving stack and evaluation harnesses, and scores cannot be replicated from weights alone without the conditions that produced them. Until independent, like-for-like comparisons and infrastructure checks appear, Beam's standing against GLM-5.2 and its other rivals stays unproven. That checking is the real launch. It starts this week.

Sources

  1. [1] Reflection, "Beam is now available", October 16, 2026 — primary source for the release, formats, partner list, Mirror CLI, first-party API beta and the self-distillation timingRead source
  2. [2] Reflection, Beam Model Card, accessed October 16, 2026 — primary source for architecture, the final benchmark table by precision, data description, safety summary and the distillation permissionRead source
  3. [3] Reflection organisation page, Hugging Face, accessed October 16, 2026 — the four Beam 1.0 repositories and their update timingRead source
  4. [4] Baseten, Beam model library page, published October 16, 2026 — licence confirmation, day-0 Model API availability, and the pricing figures exactly as displayed that dayRead source
  5. [5] Unite.AI, "Reflection Releases Beam's Open Weights Under Apache 2.0 License", October 16, 2026 — secondary coverage of the release and the partner list; the GPQA Diamond and AIME figures it attributes to Reflection's evaluations are not used as load-bearing numbers hereRead source