TL;DR
- Three large open-weight models landed inside 24 hours: Meta’s Muse Spark Open (816B total, 42B active, 1M-token context, text/image/audio/video input), InclusionAI’s Ling 3.1 Flash (560B/25B, MIT) and NVIDIA’s Nemotron 3.5 Super VL (120B) plus Bolt 74B-A9B, which fits in 24GB.
- All three are sparse mixture-of-experts: 4.5% to 12% of parameters run per token. Compute has become mid-sized. Memory has not.
- Muse Spark Open wants a 1.5 TB BF16 checkpoint across 4×8 H200 nodes, or a single B200 node for the NVFP4 build. Ling wants 8 H200s or 4 GB300s at full context. Bolt wants 24GB.
- Licences diverge as sharply as the hardware: Meta restricts commercial use, Ling is MIT, NVIDIA ships weights with datasets and recipes.
The open tier went sparse
Meta published Muse Spark Open on October 15, and the Hugging Face release write-up calls it “the most capable US open-weights model to date.” It follows Muse Glimmer from the same series and shares its tokenizer. The architecture is where the release gets interesting: 62 layers, a hidden dimension of 8192, 256 experts per MoE layer, and a router that selects 8 of them per token. (Source : Hugging Face — Welcome Muse Spark Open)
Those experts hold about 97% of the parameters — roughly 790B of the 816B. Only 8 fire per token, so each token is computed with about 42B parameters. That ratio is the story of the release: 19.4× more memory than compute per token. (Source : Meta AI Research — Muse Spark Open)
Ling 3.1 Flash pushes the same trade further (560B/25B, a 22.4× gap) and Nemotron 3.5 Super VL sits at 10×. Bolt is the outlier in the other direction: 12.2% of its weights are active, which is why 24GB is enough. Meta, meanwhile, is still unwinding the Manus acquisition that Beijing forced it to abandon — its open releases and its M&A strategy are not run by the same logic.
Compute is decoupled. The bill is not.
Sparse MoE was sold as an efficiency trick. On the compute axis, it holds. The memory axis is where builders now live.
Muse Spark Open’s grouped-query attention uses 16 key-value heads of 64 dimensions across 62 layers, so every token adds 126,976 values to the KV cache — 248 KiB per token in BF16, or about 248 GiB for a single 1M-token context, on top of the weights. The full BF16 checkpoint is roughly 1.5 TB in safetensors, which works out at 1.84 bytes per parameter. (Source : Hugging Face — Welcome Muse Spark Open)
That buys you a deployment ladder, not a download:
- 4 nodes × 8 H200 (TP=32) for BF16 — 2.5 TB of accelerator memory for a model that computes like a 42B.
- One B200 node for the NVFP4 checkpoint: 4-bit weights, what the release calls negligible accuracy loss.
- ~304 GB at 2-bit via llama.cpp, measured at 18 tokens/s on a 512GB Mac M3 Ultra.
- Two M5 Ultra Mac Studios with 512GB each, over MLX, in Meta’s own demo.
Ling 3.1 Flash draws the line elsewhere: a smaller model under MIT, but its model card still asks for eight H200s or four GB300s at full context and recommends SGLang for serving. (Source : Hugging Face — InclusionAI Ling 3.1 Flash)
The tier that actually matters: 24GB
The most consequential release may be the smallest. Bolt 74B-A9B uses native 2-bit quantisation to fit 74B into 24GB, and NVIDIA places it first among open models in its class on PinchBench and second on BrowseComp. It runs in llama.cpp, Ollama, LM Studio, Unsloth and vLLM, and shipped alongside NVIDIA’s RTX Spark laptops. A 4GB sibling follows on October 23. (Source : NVIDIA — Open From Data Center to Desktop)
Super VL — 26 trillion pretraining tokens, then roughly 3.8 million post-training samples plus RL — arrives with open datasets and recipes, and NVIDIA is selling post-training rather than inference: Scale AI tuned it for Skydance’s vendor-credit agent and EY’s due-diligence work, Mercor for coding agents on APEX-SWE, while Atlassian, CrowdStrike, Perplexity and Salesforce customise it for their own domains. (Source : NVIDIA — Open From Data Center to Desktop)
Read against July’s open-source paradox, the pattern holds: open weights are now a distribution strategy for vendors selling compute, tooling and customisation — not a gift to hobbyists.
FAQ
Can I run Muse Spark Open locally? Not the BF16 build. NVFP4 needs a single B200 node; a 2-bit llama.cpp quant ran in about 304GB at 18 tokens/s. Hosted routes — Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together — are the practical first test.
Is Ling 3.1 Flash the better deal because it is MIT? For commercial prototyping, likely yes: Meta’s release post states no licence while its developer terms restrict commercial use. But the hardware floor stays at eight H200s. The licence is cheaper than the servers.
Why does a 42B-active model need 1.5 TB of weights? All 256 experts per layer must stay resident. The router picks 8 per token; the machine holds the whole set. You pay the memory of 816B to get the speed of 42B.
Are the benchmark claims independent? No. The Muse Spark Open post carries no benchmark scores, Ling’s task claims are model-card claims, and NVIDIA’s rankings are NVIDIA’s own.
Is 1M-token context 1M usable tokens? No. The configuration exposes 1,048,576 positions and the cache cost above is real, but retrieval quality at that length is a separate measurement — as with earlier long-context releases.