Friday is mostly open weights. Meta, Black Forest Labs and Nvidia all shipped on Thursday, and the interesting numbers this time are memory footprints and license thresholds rather than benchmark charts. We close with the company most of you pay for inference, which just showed investors what its margins look like.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Muse Spark Open is free to use until you resell it at scale
Meta released Muse Spark Open on October 15 and calls it the most capable US open-weight model so far. It has 816 billion parameters with 42 billion active per token, a 1M-token context window, and takes text, images, video and audio. Weights are on Hugging Face under meta-models/Muse-Spark-Open-816B-A42B, and hosted endpoints are already up on Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together. Meta says it beats Qwen 3.8, GLM 5.3 and Kimi K3 in several categories but trails some of them on coding, per The Decoder.
Headlines said it runs on Macs. Meta's own blog says two M5 Ultra Mac Studios with 512 GB each, and the model card puts the floor at 700 GB for the NVFP4 weights and 2 TB for BF16. The license is the part to read before you ship. Any product using it has to display "Built with Muse Spark Open" somewhere visible. If you offer the model itself as an API service and your group's gross revenue passes $20 million in any 12 months, you have to notify Meta and sign a separate agreement first. Internal use and using it as a component inside your own product are carved out of that threshold. Fine-tunes and distillations count as derivatives, so the same terms follow them.
For builders
If you want to test it, use OpenRouter or Groq instead of buying hardware: run your own eval set of 20 real prompts from your product against it and against whatever you pay for now, and log cost per task alongside quality. If it wins, add the "Built with Muse Spark Open" line to your settings or about page before launch, and note in your docs whether you distilled anything from it, since the license treats those outputs as derivatives too.
Story II
FLUX 3 Image Klein gets an INT8 build that fits on an RTX 3070
Black Forest Labs released FLUX 3 Image Klein, a 5.6B-parameter open-weight model for text-to-image, editing and generation from up to ten reference images, with native 4K output. That is 38% fewer parameters than FLUX.2 [klein]'s 9B, and BFL says it gained 91 Elo points over that model in its own testing and beats Ideogram V4.5 and its own FLUX.2 [pro]. Those are internal numbers.
The hardware claim is more checkable. BFL says the INT8 build runs on GeForce RTX cards with as little as 8 GB of VRAM through ComfyUI or Diffusers, and the ComfyUI repo names the RTX 3070 and older cards without FP8 support. The Decoder reports under 9 seconds per image on an RTX 4070. Be careful with the base repo, though: its reference code peaks at 26 to 31 GB on a B200 at 1024x1024, and the optional 9B prompt upsampler adds about 28 GB more. The license is the FLUX Kommunity License v1.0, which allows commercial output for companies under $5 million in annual revenue and says that is not blanket permission for commercial fine-tuning or redistribution.
For builders
If you are paying per image for product shots or thumbnails, pull the INT8 files from black-forest-labs/flux-3-image-klein-comfy, skip the prompt upsampler on an 8 GB card, and time 50 of your real prompts. Then check your company's revenue against the $5M line before you put the outputs in anything you sell, and read the fine-tuning clause before you train a LoRA for a client.
Story III
Nvidia puts a 74B Nemotron into 24 GB and releases the recipe
Nvidia shipped Nemotron 3.5 Super VL, a 120B mixture-of-experts model with about 12B active, distilled from Nemotron 3 Ultra and built to read text, images, documents and video. Alongside it came Nemotron 3.5 Bolt 74B-A9B, which uses aggressive 2-bit weight quantization to fit in 24 GB. A 4 GB Bolt for edge devices is due later this month. Bolt runs on llama.cpp, Ollama, LM Studio and vLLM, and Nvidia is publishing the datasets and training recipes.
On Artificial Analysis's Intelligence Index, Super VL scores 26 and Bolt 21, against 58 for Claude Opus 5.5, so nobody should mistake this for a frontier model. The pitch is throughput. Nvidia cites about 630 tokens per second for Super VL with 100k input tokens, and roughly 51% on SWE-bench Pro, which it says it finishes about four times faster than competitors at slightly lower accuracy. The Decoder did not name a license, so check the model card before you deploy.
For builders
A 24 GB card like a used RTX 3090 or 4090 can now hold a 74B model. Pull Bolt in Ollama or LM Studio and give it the boring, high-volume jobs in your pipeline (classifying tickets, pulling fields out of PDFs, first-pass log triage) where a score of 21 is plenty. Keep the hard reasoning on your paid model and measure what share of calls you managed to move.
Story IV
Anthropic shows investors $1.70 back on every compute dollar
Anthropic told investors its Q3 revenue was $16.8 billion, up from $11.5 billion in Q2 and about $4.8 billion in Q1, according to the Financial Times as summarized by Simon Willison. Growth slowed from more than 140% quarter over quarter to about 45%, and The Decoder notes the Q2 spike came from Claude Code demand. The company says it earned $1.70 in revenue for every dollar spent on compute in Q3 and ran an 11% adjusted operating margin, excluding stock-based compensation. The IPO is planned for November.
The line builders should notice is distillation. People with direct knowledge told the FT that 5% to 7% of revenue came from other labs using Claude to train their own models, and that Anthropic cut this back "materially" in Q3. Willison did the math: at 5% of Q2, that is more than half a billion dollars spent distilling Claude in one quarter.
For builders
If any part of your stack generates training data from Claude outputs to fine-tune another model, reread Anthropic's usage policy today. The company is now telling investors it is shrinking that revenue on purpose, which makes enforcement more likely, not less. And if you have been holding off on a multi-year commitment because of pricing risk, positive unit economics before an IPO is about as good a signal as you will get that current pricing is not subsidized.
— The Vibe Gate news desk. We read the firehose so you can keep building.