Open-Weights Frontier: Muse Spark Open and Beam on Your Hardware

Within 48 hours this week, two US labs published frontier-scale weights — and their release pages read like two different answers to the same question: who is allowed to run this, and on what? Meta put out Muse Spark Open on October 15, billing it "the most capable US open-weight model to date." Reflection released Beam on October 16 under Apache 2.0. Both are mixture-of-experts models with million-token contexts. The interesting differences are not the benchmark tables — they're the hardware lists and the license text.

What Actually Shipped

Muse Spark Open is the open-weight release of Muse Spark: 816B total parameters, 42B active per token, natively multimodal (text, image, audio, and video in; text out), with a 1,048,576-token context window. Meta's announcement is explicit that this is a desktop story too: "the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework." The build guide specifies the rig — two Mac Studios with M5 Ultra and 512GB of unified memory each, or four with 256GB each, connected over Thunderbolt 5 with RDMA.

Beam is Reflection's first open-weight model: 501.3B total, 23.4B active, 52 layers (3 dense + 49 MoE), 1M context, text-only, knowledge cutoff June 30, 2026. It was pretrained on 23.8 trillion tokens and post-trained with large-scale asynchronous RL; Reflection positions it as competitive with GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks — a claim we report, not verify.

The Table: What Runs Where

Sources: Meta's model card, checkpoint docs, and research post; Beam's model card and the Hugging Face launch post.

Read the NVFP4 row twice. The 4-bit checkpoints are the practical entry point for both models, and both are Blackwell-native — Muse Spark Open's docs put the NVFP4 checkpoint on a single 4-GPU Blackwell node at ~537 GB, while HF's Beam post notes the ~280 GB NVFP4 checkpoint "runs natively on Blackwell" and that older GPUs like the H200 cannot. Hopper fleets still serve Beam from its FP8 or BF16 checkpoints — 16 × H200 across two nodes is the documented BF16 example — while Muse Spark Open's BF16 path means 2 TB of VRAM.

The License Difference Is the Story

Beam is Apache 2.0. Permissive in the way most software teams expect: use it commercially, ship it in products, fine-tune it, no revenue share. For a frontier-class model, those are about as loose as terms get. Ecosystem support landed fast: day-0 transformers, SGLang, and vLLM support, plus Unsloth GGUFs.

Muse Spark Open is under the Muse Model License — what Meta calls "limited commercial terms, which allow many developers to use the model for free", plus an Acceptable Use Policy. That phrasing is doing real work: "many" is not "all." The exact terms decide, and our read of the announcement pages is that the money side is not spelled out there. Treat commercial-use decisions as an open question that requires reading the license text itself, not the blog post.

Beam's Apache 2.0 versus Meta's limited-terms license is the contrast that matters here. For the wider field, this site has covered other open-weights releases — from DeepSeek's releases through Kimi K3 and GLM 5.2.

What the Active-Parameter Design Buys You

Both models exist to solve the same deployment asymmetry: serving needs the memory of a 500B-class model but computes through only tens of billions of active parameters per token. Muse Spark Open routes each token to 8 of 256 experts; Beam's router scores 256 experts and selects 8 per token. The spec-sheet lesson for capacity planning: VRAM budget tracks total parameters, throughput tracks active ones.

Speed tricks travel with both. Meta's release ships Helion kernels in SGLang with 1.03–3.95× kernel-level speedups on NVIDIA and AMD, and its dev blog shows a DSpark speculative-decoding draft config — the same technique family as DeepSeek's DSPark. Beam ships DFlash2 speculative decoding for faster generation, plus a Rust-based harness. If you've never tuned speculative decoding on a local rig, our Bonsai 27B phone-inference piece walks through running its DSpark drafter on a phone.

The Side Bet: Comfy's $25K Bounty

The third open-weights event this week is an argument, not a release. ComfyUI announced on October 15: "Higgsfield promised to open-source their entire platform. They didn't. Let's help them." — $25,000 in prizes over two weeks for the best open-source builds on the Comfy Developer Platform, backed by NVIDIA AI, RunPod, MiniMax AI, and Black Forest Labs: a $10K grand prize, RTX 5090s for Most Innovative, Most Viral, and Open Weights Only, and 10 runners-up at $250 in Comfy credits.

Read it as a signal about where leverage sits: when a vendor walks back an open-source promise, the community's counter-move is to fund the open alternative directly. It's also a concrete deadline — two weeks from October 15 — if you've been sitting on a Comfy workflow worth shipping.

Practical Starts

• Have two M5 Ultra Mac Studios? Pull the weights from Hugging Face and follow Meta's MLX path — the guide connects them over Thunderbolt 5 with RDMA.
• Have a Blackwell node? The NVFP4 checkpoints are the target for both models: ~537 GB for Muse Spark Open, ~280 GB for Beam.
• Have only Hoppers? Beam still runs, via its FP8 or BF16 checkpoints — 16 × H200 (BF16) is the documented configuration; Muse Spark Open's BF16 footprint (~2 TB) is datacenter territory.
• Just evaluating? Try Muse Spark Open on a hosted API first — seven providers carry it, including OpenRouter and Together AI — before you buy a single GPU.

What to Watch

Beam's technical report is "coming" per its model card, so its benchmark claims are vendor-stated for now. Muse Spark Open's commercial-license boundary needs the actual license text before anyone builds a business on it. And the Comfy bounty ends around October 29 — watch what the "Open Weights Only" track produces, because it will be the cleanest test of whether the bounty model moves open-source video tooling faster than vendor promises do.