Meta's Muse Spark Open Leads US Labs Storming the Open-Weights Frontier

Within roughly 48 hours, NVIDIA, Meta, and Reflection published new model weights on Hugging Face. Each company’s lead release reached the intelligence-versus-size Pareto frontier tracked by Artificial Analysis. Meta’s Muse Spark Open ranked as the highest-scoring US open-weights model on the index at publication, while NVIDIA and Reflection added smaller options for local agents and coding workloads.

The Artificial Analysis Intelligence Index combines results from several benchmarks into a single score. A model reaches the Pareto frontier when no smaller tracked model posts a higher score, making the chart useful for comparing capability against model size rather than capability alone.

“Open weights” means developers can download the model parameters and run them on their own infrastructure. Access to training data, source code, and broad commercial rights varies by release, so the three licenses differ substantially.

Muse brings 816 billion parameters to Mac Studios

Meta’s Muse Spark Open contains 816 billion parameters and activates 42 billion for each token. It accepts text, images, video, and audio, with a context window of up to one million tokens. The model card lists 62 layers, a hidden width of 8,192, and 256 experts containing about 97% of its parameters.

The sparse mixture-of-experts architecture routes each token through a subset of the network. That 42-billion active count reduces per-token computation compared with an 816-billion-parameter dense model, although the full weight set still determines storage and memory requirements.

Meta says developers can run Muse Spark Open across two M5 Ultra Mac Studios with 512GB of unified memory each or four systems with 256GB each. Its 4-bit server build requires about 700GB of GPU memory. Managed inference was available at publication through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

The custom Muse Model License permits free use for many developers under limited commercial terms. Teams planning production deployment should review its use restrictions and redistribution conditions before adopting the weights.

Nemotron squeezes agents onto smaller hardware

NVIDIA’s announcement covers three agent-focused models: Nemotron 3.5 Super VL and two smaller Bolt variants. Super VL scored 26.5 on the Artificial Analysis index, displacing Thinking Machines’ Inkling from the small-model end of the Pareto frontier.

Nemotron 3.5 Super VL has 120 billion total parameters, 12 billion active parameters, and a context window of up to one million tokens. It combines Mamba-2 sequence layers, sparse experts, attention layers, and a vision encoder. The hybrid design targets long-context agents that need to interpret documents, charts, and other visual inputs.

NVIDIA reports up to five times the single-GPU throughput of other open models in the same size category. It also claims multimodal agent accuracy comparable to its larger reasoning model at about one-fifth of the parameter count. Those results are vendor-reported and will vary with hardware, batch size, context length, runtime, and quantization. NVIDIA trained Super VL on approximately 26 trillion tokens, followed by post-training on about 3.8 million samples.

The Bolt models target local hardware. Bolt 74B-A9B activates 9 billion of its 74 billion parameters, uses native 2-bit weights, fits within 24GB of memory, and supports a 256,000-token context window. NVIDIA scheduled it for RTX Spark laptops arriving Oct. 16, 2026, followed by a 4GB Bolt E4B variant on Oct. 23. Both use the permissive OpenMDW-1.1 license, which requires preservation of the license and notices. NVIDIA also provides open datasets and training recipes.

Beam targets coding with Apache terms

NVIDIA-backed startup Reflection released weights for Beam 1.0, its first model. Beam scored 30.3 on the Artificial Analysis index, placing it between Nemotron 3.5 Super VL and Muse Spark Open.

Beam uses a sparse mixture-of-experts architecture with 501.3 billion total parameters and about 23.4 billion active during token processing. It supports a one-million-token context window, and its instruction-tuned checkpoints are available in BF16, FP8, and NVFP4 formats.

Reflection’s final model card reports 81.6% on SWE-bench Verified, which measures fixes to real software repositories; 80.1% on Terminal-Bench 2.1, which evaluates command-line tasks; and 97.7% on the 2026 AIME mathematics benchmark. Those figures apply to the BF16 checkpoint. Reflection says it pretrained Beam on 23.8 trillion tokens and then generated more than 100 million reinforcement-learning rollouts using 10,500 NVIDIA GB300 GPUs over four weeks.

Reflection also released a DFlash2 draft model for speculative decoding, a technique that proposes several tokens for Beam to verify in parallel and can improve generation speed. Beam carries the Apache 2.0 license, giving it the broadest commercial and redistribution rights among the three flagship releases.

Four models, four deployment envelopes

Deployment math can change the answer

• Benchmark the full agent loop. Composite scores and coding tests do not capture tool-call reliability, retry rates, prompt-processing latency, or performance on a team’s private repositories.
• Budget beyond the weight files. Runtime memory also covers the key-value cache, activations, serving software, and concurrent requests. Million-token prompts can add substantial prefill time and memory use.
• Test the intended precision. BF16 benchmark results do not automatically carry over to FP8, 4-bit, or 2-bit deployments. Quantization can change accuracy, throughput, and tool-use reliability.
• Verify runtime support. Mixture-of-experts routing, Mamba-2 layers, multimodal encoders, and multi-system Apple deployments require compatible kernels and serving frameworks.
• Review license compatibility. Apache 2.0 offers the clearest path for modification and redistribution. OpenMDW-1.1 and the Muse license impose different notice and commercial-use conditions.

US labs fill the efficiency curve

For much of 2026, Chinese labs including DeepSeek, Alibaba, Z.ai, StepFun, and Xiaomi occupied the top of the open-weights intelligence chart. MiMo-V2.6-Pro and DeepSeek V4.1 Flash remained the overall leaders at publication, while US releases were scarce near the frontier.

The new models give developers US-built options across several hardware tiers, from a 4GB edge model and a 24GB consumer-GPU release to multi-system Apple deployments and Blackwell GPU clusters. Model quality, runtime support, memory requirements, and licensing still differ enough to require workload-specific testing, but teams now have more viable candidates at each point on the efficiency curve.