Reflection Released Beam's Weights 11 Days After Announcing It. Meta's Bigger Muse Spark Open Came a Day Earlier With an Access Form and a $20 Million Line.

Ten days ago we ran a piece with a blunt headline: Mistral and Reflection had announced open-weight models a day apart, and neither had released the weights. Half of that is now out of date.

On Friday, October 16, 2026, Reflection made Beam available, with weights on Hugging Face under Apache 2.0. That is 11 days after the Monday, October 5 announcement that promised the weights "later this month." The earliest commit in the main repository, titled "Beam is available," is stamped 10:34 UTC.

It did not land alone. A day earlier, on Thursday, October 15, Meta Superintelligence Labs released Muse Spark Open, a larger model with a stronger benchmark table. So the US now has two serious open-weight releases in two days. They are not open in the same way, and that difference is the story.

The Two Releases Side by Side

Both are sparse mixture-of-experts models with 256 experts and a 1 million token context. Beyond that, they split on almost every line that matters to someone deciding what to download.

The license row is the one we would read twice. Beam is Apache 2.0 and anyone can pull it from Hugging Face without asking. Muse Spark Open ships under a Meta license that grants broad rights to use, modify, fine-tune, distill and redistribute, then adds two conditions. Products must display "Built with Muse Spark Open" somewhere a user or reader would see it. And any company selling model access through an API whose group revenue tops $20 million in a 12-month window must notify Meta and sign a separate agreement first.

The Hugging Face gate sits on top of that. To download the Meta weights you submit your name, date of birth, country, affiliation and job title, and the repository is set to manual approval. The weights are downloadable, but you have to ask first. That is not what most developers picture when they hear "open."

Meta Has the Stronger Table

None of this means Beam is the better model. Meta's model card does something we like: it puts Beam and Mistral Large 4 in its comparison columns, next to the Chinese leaders. On Meta's own numbers, Muse Spark Open wins most of the head-to-heads with Beam.

Beam keeps two rows: Terminal-Bench 2.1 and GPQA Diamond. On the agentic coding test we watch most, DeepSWE, Meta lists Muse Spark Open at 60 against Beam's 44. Kimi K3 still sits above both on most lines, which matches what Reflection conceded on day one: "frontier open models like Kimi K3 remain ahead on raw capability."

Reflection's pitch was never raw score. It was efficiency, and the 23 billion active parameters are the evidence. Beam activates about 4.6 percent of its weights per token; Muse Spark Open activates about 5.1 percent, but of a model 63 percent larger. Reflection's card claims comparable solution quality with roughly 3x fewer reasoning tokens. VentureBeat's reading of Reflection's charts puts Beam at roughly 25,000 tokens per Terminal-Bench 2.1 attempt against roughly 80,000 for GLM-5.2. If that holds up in independent runs, cost per finished task could favor Beam even where its score trails.

The Numbers Moved Between Launch and Release

One detail worth flagging. Reflection's October 5 post now carries the note "Update 10/8/26: We've updated Beam's results," and lists Beam at 44.4 on DeepSWE v1.1. The model card published with the weights lists 43.3 for the BF16 checkpoint, 43.0 for FP8 and 42.7 for NVFP4. The gaps are small. But it is a reminder that until today, every Beam number came from a model nobody outside an early access group could run.

Vendors also disagree about their rivals. Reflection's October 5 post, as updated, puts Kimi K3 at 88.3 on Terminal-Bench 2.1. Meta's card puts the same model at 85.0. Neither is wrong on its face, since harnesses and settings differ, but it is why we treat any cross-vendor table, including ours above, as a sketch.

What It Takes to Run Each One

Hardware is where Beam's smaller footprint pays off. Hugging Face's launch post puts the BF16 checkpoint at about 1 TB and NVFP4 at about 280 GB, and calls four B200s the practical setup. Meta says Muse Spark Open's BF16 weights are about 1.6 TB and NVFP4 about 530 GB, enough for four Blackwell Ultra GPUs or eight H200s. Meta also optimized it to run across two 512 GB M5 Ultra Mac Studios using MLX, which is a genuinely useful option for a small team with no data center.

Distribution looks similar on both sides. Meta lists Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry for weights, and seven hosted inference providers. Reflection says it launched with "over 40 partners and integrations," naming AWS, Google Cloud, Microsoft Azure, Databricks, Ollama, OpenRouter and Together AI, with Dell and HPE offering on-premises and air-gapped installs. It also released Mirror CLI, the coding harness it says Beam was trained with.

Mistral Is Now the One Waiting

Mistral Large 4 remains API-only. Its October 6 announcement says "We will release the weights by the end of the month," and Quartz reported the company named Tuesday, October 27 as the date. When we checked Mistral's Hugging Face organization on Friday afternoon UTC, no public Large 4 repository was listed, and the announcement still names no license. The irony is that Meta's card already lists Large 4 at 62 on DeepSWE, above Muse Spark Open, and at 38 on the Artificial Analysis Intelligence Index, one point under Meta's 39. The best scoring of the three on that coding row is the only one you cannot download.

Caveats

Every capability figure here is self-reported by the lab that trained the model, or by a rival lab in its own comparison table. Reflection's 3x token claim and VentureBeat's token counts come from Reflection's charts, not an independent run. The weights went public this week, so the independent numbers are only starting to arrive. Meta's safety results, including a 99.2 percent jailbreak refusal rate on an internal cybersecurity benchmark, come with Meta's own warning that they "should not be assumed to apply to modified models."

Our Take

For two years the argument about American open weights was whether anyone would ship at frontier scale. This week two labs did, and the argument moved on to what "open" buys you.

On raw capability, Meta wins this round, and we would not pretend otherwise. If you want the strongest US model you can host yourself, Muse Spark Open is it on the evidence published so far. But for a startup building an inference business or a product it hopes to sell, Beam is the safer foundation. Apache 2.0 has no revenue line, no attribution banner and no approval queue. A $20 million threshold sounds generous until your company is the one approaching it, and then your foundation model's owner is also your counterparty.

We also give Reflection credit for doing what it said. It promised the weights later in October, and it delivered with two weeks of the month to spare, alongside FP8 and NVFP4 builds, a drafter model and a safety report. Mistral now has 11 days to do the same.

Three signposts for the next 60 days: whether independent evaluators confirm Beam's token efficiency on cost per finished task, which VentureBeat says it is now testing; whether Mistral Large 4's weights arrive by October 27 and under what license; and whether a hosted price for Beam appears on OpenRouter at or below Muse Spark Open's $1.25 and $4.25.

Sources: Reflection (Beam is available), Reflection (Introducing Beam), Beam model card, Hugging Face blog, Meta (Build with Muse Spark Open), Muse Spark Open model card, Mistral, VentureBeat, Unite.AI, Quartz via Yahoo Tech, OpenRouter.