Beam 1.0: Reflection Open-Sourced the Weights, and Artificial Analysis Scored It 30

Two things happened to Reflection's Beam 1.0 this week, and they point in opposite directions. The model — Beam 1.0-501B-A23B, a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active per token — is no longer a promise: the weights are on Hugging Face under Apache 2.0, the 273 GiB NVFP4 build is downloadable now, and there is a first-party API with a documented contract and a real rate-limit page. At the same time, the first independent number anyone has published for it is 30 on the Artificial Analysis Intelligence Index — nineteenth of 120 models evaluated on index version v4.3.2, and above the median of 19 that Artificial Analysis records for comparable open-weight models of similar size. Released vendors usually get to choose which of those two facts leads. Here, both are true at once, and the honest read is that Beam 1.0 arrived cheaper than expected but scored lower than its own launch post implied.

This is a follow-up to our earlier coverage, and the framing is deliberate: Beam 1.0 is not new. Reflection announced it on 5 October 2026 and updated its benchmark tables on 8 October. What changed inside the last seven days is everything a buyer actually needs — weights, hosting, an API, limits, and outside measurement. If you read the launch coverage and decided to wait, the reason to wait is gone.

What actually shipped this week

Reflection's "Beam is now available" post is dated 16 October 2026 and it closes the gap that made the launch post so hedged. On 5 October the weights were something Reflection said it "will release." As of this week they exist, in three precisions, plus a speculative-decoding drafter:

• BF16 — 501.3B parameters across 188 safetensors shards, uploaded to the reflection/Beam-1.0-501B-A23B repository on 7 October and revised since.

• FP8 — same 501.3B parameter count in E4M3, 468.72 GiB on disk.

• NVFP4 — a packed 4-bit build at 273 GiB, which is the one Reflection's own serving guides tell you to download.

• DFlash2 — a 1.93B-parameter speculative decoding drafter, uploaded 13 October, published as a separate repo and documented as its own serving optimisation.

All four carry the apache-2.0 tag, and the licence file is in the repository rather than in a blog sentence, which is the difference between a licence and a promise. Reflection also states it does not restrict distillation, fine-tuning, quantisation or merging.

The serving story is equally concrete. There is day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and N​VIDIA Dynamo, and Reflection's docs carry step-by-step guides for SGLang, vLLM, TokenSpeed and a Helm chart for Kubernetes. The reference self-hosting requirement is four N​VIDIA B300 SXM GPUs on NVLink — a single box, not a cluster, which is a much lower bar than a 500B model normally implies and follows directly from the 23B activation ratio.

There is also a verification tool. The open-source beam-verifier CLI runs a conformance suite against any live Beam endpoint and then evaluates it on the main benchmarks so you can compare your deployment, or a provider's, against Reflection's reference scores. Anything that lets a buyer check a vendor's numbers rather than trust them is worth knowing about.

The first outside score, and it is not the launch-post number

Artificial Analysis lists Beam 1.0 with an Intelligence Index of 30, ranked #19 of 120. The same entry shows the model used 79M output tokens across the Index — "#10 / 46", against a median of 130M — which is the token-efficiency claim showing up in an independent harness for the first time. Artificial Analysis also shows no pricing and no speed for Beam, because as of this writing it still records "No publicly available API" for the model, pointing only at self-hosting. That gap is itself information: the API exists, but not yet in a form the measurement firms have picked up.

Read the 30 carefully. It is not a verdict on the vendor's benchmark table — the two are measuring different things on different harnesses. Reflection's model card, published with the weights, reports 81.6% on SWEBench Verified, 80.1% on Terminal Bench v2.1 and 90.6% on GPQA Diamond for the BF16 checkpoint, and the launch post is explicit that frontier models such as Kimi K3 remain ahead on raw capability, with Beam's argument being "efficiency at inference time" rather than parity at the top. Those figures remain vendor-reported and, as far as we can find, nobody has reproduced them yet. But the release of a public endpoint plus the verifier CLI is what makes reproduction possible at all, and a 30 on a 120-model index is the first number in this story that Reflection did not produce.

The price, the limits, and what they mean per task

The API side is now documented rather than waitlisted. The endpoint is https://api.reflection.ai/openai/v1, O​penAI-compatible for Chat Completions and Models, with a native base URL at api.reflection.ai/v1 if you want Reflection's own shape. Model ID: Beam-1.0-501B-A23B. The docs give the served limits precisely, and they are narrower than the 1M-token training figure that circulated at launch:

• Context window — 1,048,576 tokens, with a maximum output of 131,072 tokens and a default max_tokens of 32,768.

• Reasoning — mandatory, not optional. Supported efforts are low, medium, high, xhigh and max, with max as the default. Reasoning tokens count toward your rate limits.

• Capabilities — reasoning, tool calling and structured outputs, text in and text out.

• Rate limits — measured per organisation in requests and tokens per minute and per UTC day, with concurrent-request caps and an optional daily token allowance. Overruns return 429 with either rate_limit_exceeded or daily_allowance_exhausted, and when the daily allowance is spent the response sets x-should-retry: false, so naive retry loops will simply spin.

What the docs do not carry is a public per-million-token price. Reflection's announcement says only that it is "free to start," and neither the docs nor the announcement publishes a rate card. That is the one line still missing from the picture, and it is the line that decides whether the efficiency argument means anything on a bill. A model that reasons at every request, defaults to maximum effort, and claims 3–4× lower inference compute than GLM 5.2 on advanced reasoning is a different proposition at $0.50 per million output tokens than at $5.00 — and nobody outside Reflection can currently say which it is.

Two practical cautions from the documentation worth carrying into a pilot. First, the default effort is max, so the out-of-the-box configuration is the most expensive one; a cost-conscious integration should set reasoning_effort explicitly from the first call. Second, limits are per organisation and shared across projects and keys — adding keys does not raise them, which is the kind of detail that only bites after you have already built the second integration.

Where it sits against the models you would actually compare it to

The interesting comparison is not against frontier flagships, where Reflection concedes the point, but against the open-weight models Beam is priced and sized against. On our own catalogue today:

• GLM 5.2 — 1M context, $1.40 per million input tokens and $4.40 per million output, fully served.

• Kimi K3 — 1,048,576-token context, $3.00/$15.00, and the model Reflection's own table puts ahead on Terminal Bench.

• Qwen 3.8-Max — 1M context, $2.00/$6.00, the 2T-plus class Beam claims a larger efficiency advantage over.

Beam's pitch is that it gets near GLM 5.2 on advanced reasoning for a fraction of the inference compute. If that holds under an independent harness, it is worth real money at high request volume. The 79M-token verbosity figure is the first evidence that it might. The missing price is why you cannot finish the arithmetic yet.

This is also the situation where routing removes most of the risk of being wrong. Beam is not on our catalogue at the time of writing, so we cannot serve it for you — but every model above is, behind one OpenAI-compatible key with the provider's list price passed through and no markup, so a vendor price cut reaches your bill the same day rather than at the next contract. The practical move is to keep your production path on a model with a published rate card and put Beam's public beta behind a routing rule, so an unproven first-party API never becomes a single point of failure. Automatic failover does the rest: if the beta rate-limits you with a 429 at 02:00, the request lands somewhere that will not.

Who should move, and on what evidence

If your constraint is open weights and a permissive licence, this week is the week the answer became yes rather than soon. Apache 2.0, three precisions, a 273 GiB build that fits on one four-GPU B300 node, day-zero vLLM and SGLang support, a verifier to check your own deployment, and no restriction on fine-tuning or distillation. For anyone whose blocker was "the weights are not out," that blocker is gone, and self-hosting removes the API price question entirely.

If your constraint is a production dependency on a hosted endpoint, be more careful. The API is real and documented, but it launched this week, it enforces organisation-wide limits that a second key will not raise, it reasons at maximum effort by default, and its price is unpublished. Run the conformance suite from the verifier against it before you trust it with anything that has a deadline.

If your constraint is capability at the frontier, nothing this week changes your answer. A 30 on the Intelligence Index, a vendor table that concedes ground to Kimi K3 on Terminal Bench and to GLM 5.2 on AIME, and no independent reproduction of the headline numbers add up to a model worth trying, not a model worth betting on.

The one thing to watch next is the price. Reflection has shipped the hard half of a launch — weights, serving stacks, docs and a public endpoint — and left the easy half, a rate card, unpublished. When that number lands, the 3–4× efficiency claim becomes a spreadsheet instead of a slogan, and the comparison against GLM 5.2 or Kimi K3 becomes a decision rather than an argument about methodology.

A result like this is also a reminder that model mistakes are already documented: the AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.