This edition covers 7:38 am on 15 October to 7:38 am on 16 October 2026, Sydney time. It records three verified model releases, focusing on weight access, local deployment conditions and the limits of the evidence.
Meta releases Muse Spark Open weights
On 15 October, Meta released the multimodal Muse Spark Open with downloadable weights and hosted APIs. Its stated specifications are 816 billion total parameters, 42 billion active parameters and a one-million-token context window. Unlike the proprietary Muse Spark 1.3 service launch on 2 September previously recorded in this series, the new development is access to Spark Open weights. This does not establish that the two checkpoints are identical.
Apple published an MLX distributed-inference guide the same day, demonstrating configurations with two or four M5 Ultra Mac Studios. Its four-machine test reported more than 36 generated tokens per second at an 8,192-token context. This is developer testing, not a result for one machine or every context length. The Verge independently reported the release and collaboration, but did not reproduce the performance results.
The weights use the Muse Model License with limited commercial terms, rather than unrestricted commercial permission. Meta says a full safety evaluation report is still forthcoming. Direct references are the Meta and Apple publications and The Verge report listed below.
NVIDIA releases Nemotron 3.5 Super VL and Bolt 74B-A9B
NVIDIA made Super VL and Bolt 74B-A9B available on 15 October. Bolt E4B, introduced alongside them, remains scheduled for 23 October and is not counted as delivered. Super VL combines visual understanding with reasoning and tool use. It has 120 billion total parameters and activates 12 billion per token. The release includes base, post-trained and NVFP4 checkpoints, training recipes and data.
The NVFP4 recipe uses four-bit weights and activations on Blackwell, and four-bit weights with 16-bit activations on other platforms. NVIDIA says Bolt 74B-A9B fits in 24 GB of memory using its native two-bit configuration. This is not a memory guarantee for arbitrary context lengths or concurrency.
Independent evaluator Artificial Analysis reports Intelligence Index scores of 26.5 for Super VL and 21.4 for Bolt. These are results within a particular evaluation framework, not universal capability judgements. Its Super VL speed test used a pre-release DeepInfra NVFP4 endpoint and should not be treated as a local-hardware result. Direct references are the two NVIDIA publications and the Artificial Analysis evaluation.
FLUX 3 Image Klein releases base and distilled weights
Black Forest Labs released FLUX 3 Image Klein weights and API access on 15 October. The model generates and edits images using up to ten reference images. Its 5.6-billion-parameter figure covers the image generator, excluding the separate text encoder. The base version is intended for fine-tuning, and the distilled version for generation and editing. VentureBeat also reported the release.
The official model card says bounding boxes guide placement and scale rather than acting as hard pixel masks; edits can affect content outside a box. Quantised file sizes are not total inference-memory requirements, so the launch page's low-memory configuration is not presented as a uniform requirement for every version. Claimed quality advantages remain developer evaluations; this review did not obtain a complete independent reproduction.
Hugging Face downloads require acceptance of conditions and sharing contact information. The model card describes a FLUX Kommunity License exception for commercial output generation by qualifying users with gross annualised revenue below US$5 million, including affiliates. It is not blanket permission for commercial fine-tuning or distribution. Direct references are BFL's announcement, model card and the VentureBeat report.
Sources
Meta AI Research; Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date; 15 October 2026.
https://research.meta.ai/blog/muse-spark-open
Apple Machine Learning Research; Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio; 15 October 2026.
https://machinelearning.apple.com/research/mlx-distributed-inference
The Verge; Meta and Apple team up to make Muse AI models run better on Macs; 15 October 2026.
https://www.theverge.com/news/1012293/meta-apple-muse-spark-open-ai-model-mlx
NVIDIA; Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models; 15 October 2026.
https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/
NVIDIA Technical Blog; NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning; 15 October 2026.
https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/
Artificial Analysis; NVIDIA Releases Nemotron 3.5 Super and Nemotron 3.5 Bolt; 15 October 2026.
https://artificialanalysis.ai/articles/nemotron-3-5-super-and-bolt
Black Forest Labs; FLUX 3 Image Klein; 15 October 2026.
https://bfl.ai/blog/flux-3-image-klein
Black Forest Labs; FLUX 3 Image Klein Base model card; supplied with the 15 October 2026 release, checked on 16 October 2026.
https://huggingface.co/black-forest-labs/flux-3-image-klein-base
VentureBeat; Black Forest Labs opens FLUX 3 Image Klein weights, with free commercial use limited to smaller businesses; 15 October 2026.
https://venturebeat.com/technology/black-forest-labs-opens-flux-3-image-klein-weights-with-free-commercial-use-limited-to-smaller-businesses
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.