Beam 1.0 vs Qwen 3.8: A Delivered Weight Release Against a Price List

On 5 October 2026, Reflection Beam 1.0 was announced as an open-weight model whose weights were not out. On 16 October 2026 they shipped — four repositories on Hugging Face under Apache 2.0, in BF16, FP8 and NVFP4 precisions, plus a 2B speculative-decoding drafter. Qwen 3.8 has been the opposite story for two and a half months: A​libaba's generation went live on 3 August 2026 with Qwen3.8-Max as a priced API at $2.00 per million input tokens and $6.00 per million output, one open-weight sibling following under a custom licence and a dense 27-billion-parameter model under Apache 2.0. Both models claim a one-million-token context. Both reason, both call tools, both are text-only at the interface. The difference worth understanding is which one you can budget for and which one you can hold, because after 16 October each model has exactly one of those and not the other.

Both figures above are the vendor's, and both models have now been measured by Artificial Analysis — which is rare enough in this field to be the reason this comparison can be written at all.

The state of play, model by model

• Parameters — Beam 1.0 is a sparse mixture-of-experts model at 501.3B total with 23.4B active per token, across 52 layers, three dense and 49 MoE, with 256 experts in the configuration the card describes. Qwen3.8-Max's parameter count has never been disclosed by A​libaba; the open sibling of its core, Qwen3.8-2.4T-A95B, is a 2.4-trillion-parameter MoE with roughly 95 billion active, which is as close as anyone outside the company can get to the flagship's size.

• Context window — 1,000,000 tokens on Beam 1.0's model card. 1,000,000 tokens on OrcaRouter's listing for qwen/qwen3.8-max, and 984,000 on Artificial Analysis's page for the same model. The two are the same figure rounded differently and are not a point of difference between the models.

• Modality — Beam 1.0 is text in, text out, stated plainly on its card. Qwen3.8-Max accepts text and images and returns text, which is a real capability difference rather than a formatting one.

• Licence — Beam 1.0 is Apache 2.0, with no user caps and no separate commercial tier. The Qwen 3.8 generation is mixed: the 2.4T Max core checkpoint is downloadable under a custom licence, the dense 27B under Apache 2.0, and the flagship Max itself has no weights at all.

• Price — not published for Beam 1.0 anywhere on Reflection's own properties, and recorded as unavailable on Artificial Analysis's page for the model. $2.00 and $6.00 per million tokens for Qwen3.8-Max, published by A​libaba and listed on our own route at the same figures.

• Independent score — Artificial Analysis Intelligence Index v4.3.2 puts Beam 1.0 at 30, ranked 19th of 120 in its comparison class. The same index has Qwen3.8-Max (0902) at 45, ranked 33rd of 217, and AA has evaluated the Max line at more than one build with the number moving between them.

• Speed — not available for Beam 1.0; the page records output speed as unknown, because there is no public endpoint to measure. 38.8 output tokens per second for Qwen3.8-Max, which AA describes as notably slow for its tier.

What the sixteen days between announcement and release actually changed

The gap is worth describing precisely, because it is the sort of thing that gets flattened once the weights are out.

Reflection's launch post said the weights, technical report, model card and developer artefacts would follow "later this month," and promised Apache 2.0. Between 5 October and 16 October, the only thing a reader could do with Beam 1.0 was join a waitlist. The model had a benchmark table, a published training account — 23.8 trillion tokens on 6,144 GB300 NVL72 GPUs in under four weeks, an RL stage on 10,500 GB300s, more than 100 million rollouts — and no download.

Then on 16 October three quantisation variants and a drafter appeared in a Hugging Face collection, and the card arrived with a full evaluation table across all three precisions. For a reader deciding what to run, the important consequence is not the benchmark numbers, which are still Reflection's own and unreproduced. It is that the model moved from the category "announced" into the category "yours to inspect, quantise, fine-tune and serve." A licence you can act on is a different asset from a licence you have been promised.

Qwen 3.8 took the opposite route to the same place. A​libaba opened the generation with a priced, served flagship on 3 August 2026 and released weights in stages afterwards — the 2.4T core on 12 August, the dense 27B on 14 August, the Flash-Next preview later that month. Nobody had to wait on a promise, because the thing you could buy came first and the thing you could download came second.

The two pricing situations are not symmetrical

Qwen3.8-Max is one of our routes. It sits at $2.00 per million input tokens and $6.00 per million output with a one-million-token context, at the provider's list price passed through with no markup — so if A​libaba moves that rate, your bill moves the same day rather than at the next contract renewal. That is what "the model has a price" means in practice, and it is the reason this article can put a worked number in front of a reader for one side of the comparison and not the other.

Beam 1.0 has no such number. Twelve days after the announcement it was a waitlisted preview with a sign-up wall; on 16 October it became a download. Neither state has a rate attached to it. Any dollar figure you see quoted for Beam 1.0 is somebody's hardware estimate, and the card supplies no serving recommendation, no hardware floor and no throughput figure to estimate from. The price row is the single largest hole in the model's documentation, and it is a hole that a 501B parameter count makes expensive to fill.

There is a version of this comparison where the sizes make the cost argument for you — 23.4B active parameters against a rumoured 95B is a four-to-one ratio in compute per token, and if the smaller model were priced it might well be the cheaper one to run. That argument cannot be made here, because one side of it has no rate. What can be said is that a model with a published price and a model with a published licence are both, in different senses, ready to use, and a buyer needs both rows to choose.

Capability, with the sourcing attached

Beam 1.0's card reports SWE-bench Verified 81.6% on the BF16 checkpoint, Terminal Bench v2.1 at 80.1, SWE Bench Pro v1 at 66.1, DeepSWE v1.1 at 43.3, GPQA Diamond at 90.6, AIME 2026 at 97.7, LiveCodeBench at 88.3, MCP Atlas at 80.8, LongBench v2 at 67.5 and IFBench at 79.6, with the FP8 and NVFP4 numbers typically a point or two lower. Reflection's own positioning is careful: it describes Beam as competitive with larger open models and approaching Qwen 3.8-Max, matching GLM-5.2 on reasoning while using three to four times less inference compute, and acknowledges that Kimi K3 still leads on raw capability. All of it is vendor-reported and none of it has been reproduced by a third party.

Qwen3.8-Max's numbers come from two directions. A​libaba reports its own, including 86.6% on Terminal Bench 2.1, and the enterprise pricing sheets that go with a flagship launch. Artificial Analysis's independent composite is the more useful row: 45 on Intelligence Index v4.3.2, which places the model well above the median for its comparison class of 27 but below the frontier, and a measured output speed of 38.8 tokens per second that AA describes as slow even against a tier median near 88. A flagship that is 15 points ahead of Beam 1.0 on the same independent index and costs $2.00 to talk to is a different proposition from one that is ahead on a vendor's own table.

The one number that would settle the efficiency argument — tokens per unit of energy or per dollar of compute, measured on identical hardware — does not exist for either model. Reflection's supporting chart approximates FLOPs as roughly twice the active parameter count times generated tokens and says in its own caption that this excludes prefill, attention and serving overhead. That is an honest disclosure of a limit, and it is still not a measurement.

What each one is for

Reach for Qwen3.8-Max when you need the thing to be callable and accountable. There is a rate, a context figure and an independent score, it reads images as well as text, and it answers on a key you can hold today. The costs are honest ones: it is expensive, at $2.00 and $6.00 per million tokens, and the independent measurement says it is slow for its class, so a high-volume workload will notice both. The open weights in the generation exist, but the checkpoint that backs the flagship name is under a custom licence and datacenter-scale, so self-hosting the exact thing you are buying is not the escape hatch it sounds like.

Reach for Beam 1.0 when you want to own the stack. Apache 2.0 with three quantisation levels, a 1M-token context, an always-on reasoning budget you tune rather than toggle, and an evaluation table pointed at coding and agentic work. The price of that ownership is everything around it: a serving bill you have to estimate yourself, a parameter count that makes the estimate large, and no independent score to check the vendor's table against — the Intelligence Index figure of 30 is real and third-party, but it is one composite, not a per-benchmark verification.

And for the mixed case — a team that wants the open weights eventually but needs an endpoint this quarter — the ordinary answer is both: route the work through a served model at a published rate while the self-hosted deployment gets built and benchmarked on your own tasks. OrcaRouter runs more than 200 models behind one OpenAI-compatible key with automatic failover across providers, so a self-hosted build can take a fraction of traffic and hand back the rest when it degrades, and no second contract is required to try a model you have not committed to.

What to watch next

One thing changes this page and it is a single row. If Reflection or a serving partner publishes a Beam 1.0 rate — for the API, or for a hosted FP8 or NVFP4 deployment — then the comparison becomes arithmetic: 23.4B active parameters against a rumoured 95B, two independent index scores, and two prices. That is a page worth reading. Until then the truthful summary is that Reflection delivered the half it had promised, A​libaba delivered the half it had priced, and both models are waiting on the other half.

That gap between what a scoreboard measures and what a deployment meets is the reason the record matters: the documented AI failures and agent incidents around 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.