Reflection Beam vs Muse Spark Open: Which Open Model Fits

For much of 2026 the open-weight models at the top of published benchmark tables, such as Kimi K3 and GLM 5.3, came from Chinese labs. In the space of two days, two US labs answered. Meta released Muse Spark Open on October 15, an 816-billion-parameter model that takes text, images, video and audio, under its own Muse Model License. Reflection released Beam on October 16, a 501-billion-parameter text-only model built for coding and agents, under Apache 2.0. Both offer a one-million-token context. Almost everything else about them points in a different direction.

We covered each launch on its own: Muse Spark Open's weights, licence and hardware, and Beam's licence, hardware and access. This post puts them side by side for the decision most teams actually face: which one to trial for a self-hosted or resold agent model.

One rule governs everything below. Each vendor ran its own benchmarks, with its own harnesses and settings. An independent measure of both now exists: Artificial Analysis's Intelligence Index, published October 16, scores Muse Spark Open at 39 and Beam at 30. Where the two cards report the same benchmark we show the vendor figures together, labelled, and say what they cannot tell you.

Both are mixture-of-experts models, which split most layers into many smaller expert networks and run only a few for each token. That is why each has two parameter counts: a total that sets the memory you need, and an active count that sets the compute per token. The table uses each vendor's own figures.

Read across the rows and a pattern appears. Muse is the larger, more general model, built to see and hear as well as read. Beam is the leaner specialist, built for text, code and tool use, and released under Apache 2.0, a standard permissive licence. Neither is a cut-down version of the other; they make different trade-offs for overlapping jobs.

For most commercial teams the licence settles the question before benchmarks matter. Beam's Apache 2.0 permits commercial use, modification and redistribution, and Reflection's model card says it does not prohibit distillation or training other models on Beam's outputs. There is no revenue level at which the terms change and no notice your users must see.

Muse's licence, as read from Meta's model card on October 16, is friendlier than its headline suggests but has two triggers. Every product built on the weights must prominently display "Built with Muse Spark Open". Anyone offering the model as a service with more than $20 million in aggregate gross revenue over any twelve months must notify Meta and sign a separate agreement; internal use and research are exempt. The licence also says the model is not intended for users under 18, and attaches a usage policy.

The failure case is an agency one. A studio builds a client's assistant on Muse, ships it under the client's brand and forgets the attribution line, which has no revenue threshold. The same project on Beam has no line to forget. Our audit of 2026 open-weight licences sets both against the other major releases.

Both vendors name four Blackwell Ultra GPUs for their 4-bit builds: Reflection's serving guides assume four NVIDIA B300s, which are Blackwell Ultra parts, and Meta's launch post names four Blackwell Ultra GPUs or eight H200s. The difference is how much of that memory the weights leave free. Beam's NVFP4 download is 273 GiB; Muse's is about 530 GB, and Meta's model card asks for at least 700 GB of VRAM.

Long conversations widen the gap. Hugging Face's write-ups estimate Muse's attention cache at about 250 KB per token, roughly 266 GB at the full million tokens, which by our arithmetic is about 32 GB at 128,000 tokens. For Beam, which uses sliding-window attention on most layers, Hugging Face calculates about 4.6 GiB at 128,000 tokens. On the same four GPUs, Beam leaves far more room for concurrent users and long agent sessions.

Muse has one hardware path Beam lacks: Meta describes running the 4-bit build through MLX on two M5 Ultra Mac Studios with 512 GB each, which our M5 Ultra Mac Studio economics post prices against hosted API tokens. The failure case on either model is sizing for the weights alone and finding that one long session consumes the memory you thought was spare.

Six benchmarks line up across the two cards. Reflection ran Beam at temperature 0.7 and top-p 0.99, at an effort level its card does not state, and reports the BF16 column below. Meta ran Muse at xhigh effort, naming its preferred harnesses on some rows. Same name, different runs: read the table as two vendors' self-reports placed next to each other, not as a contest.

Beam leads narrowly on two rows, Terminal-Bench 2.1 and GPQA Diamond. Muse leads on four, by wider margins on MCP-Atlas, HLE with tools and SciCode, and by nearly seventeen points on DeepSWE. That last gap is large enough to be worth testing directly if your work is repository-scale software engineering, and small enough in harness terms (Meta used its preferred setup) that a team should not treat it as settled.

Two more rows are left out. Beam's card reports HLE without tools; Muse's reports HLE-Diamond without tools, a different subset. On SWE-Atlas QnA, Reflection's final card gives Beam 31.8% while Meta's table lists it at 35 beside Muse's 58, so we could not settle which Beam figure to use. Meta's table does include a Beam column, but its figures match the preview scores Reflection published on October 5 (78.7 on MCP-Atlas, for example, against 80.8 on the final card); the table above uses Reflection's final card.

This is the second hard filter after the licence. Muse Spark Open accepts text, images, video and audio, and Meta says it was trained on data from more than 100 languages. Beam takes text in and text out, and Reflection says it was developed predominantly for English. There is no setting that closes that gap.

The jobs it decides are concrete. A browser agent that reads screenshots, a support agent that receives photos of a damaged product, or a pipeline that extracts data from scanned invoices: each needs Muse, or a different model entirely; an assistant for a Spanish-language market is also safer on Muse. A coding agent working on a repository, a document summariser fed clean text, or a tool-calling back-office agent can use either.

Meta's line, published the day before Beam's weights were released, is a claim about capability on a desktop-class setup, and the qualifier matters: Beam names no Apple silicon path at all. For data-centre deployment, where both run, the question is not which is "most capable" but which inputs and licence your product needs.

Neither lab publishes a price for its own API: Reflection's first-party API is a free public beta, and Meta's launch post lists no Meta endpoint for Muse Spark Open. OpenRouter lists both models' host prices side by side. On October 19, Muse's route listed $1.25 per million input tokens and $4.25 per million output. Beam's four providers ranged from $0.25 input and $1.00 output (Nebius, FP8) to $1.30 and $4.30 (Baseten, FP4). The headline route, CoreWeave's FP4 build, listed $1.00 and $3.00 on October 19 and had fallen to $0.30 and $1.00 by October 20.

Defaults complicate any comparison. Hugging Face gives Beam's default effort as medium, while OpenRouter's Beam route and Reflection's own API reference default to max. Meta's card defaults Muse to xhigh, while OpenRouter's Muse route defaults to medium. A team that compares the two through OpenRouter at default settings is comparing Beam at its highest effort with Muse near its lowest, and will draw the wrong conclusion about both cost and quality.

The fix is simple and rarely done: pick one effort level per job, set it explicitly on both models, and compare cost per finished task, not price per token. Our comparison of Muse with Kimi K3 and GLM 5.3 sets out the effort settings behind each of Meta's published figures.

The routing below is ours, built from the licences, the input types and the vendor figures above. It is a starting point for a trial, not a verdict from one.

Three gaps remain. Apart from Artificial Analysis's composite index, no one has published an independent run of both models on these benchmarks, so the benchmark comparison rests on self-reports. Each vendor reports safety on its own scale: Meta's card reports a very low score on StrongREJECT, an attack benchmark where lower is safer, and Reflection reports refusal and compliance rates on its own Iris suite. Those numbers measure different things and cannot be lined up. And neither card says how either model behaves after weeks inside a real agent loop, which is where most of the cost and most of the failures live.

That is the case for a short, structured trial rather than a choice from the tables. Take twenty real tasks, run both at the same effort on the same provider type, record tokens, time and outcome, and keep the configuration beside the result. If you want help running that trial and wiring the winner into your agents, our AI transformation team does exactly this.