Wednesday has something for every layer of the stack. Apple finally lets any iOS app talk to Siri AI, Anthropic published a way to prove which model actually produced a reply, Soniox shipped a speech model that knows what your voice agent just said, and Tencent put out a small open model that does detection, OCR and matting in one pass.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Siri AI is open to every app in the iOS 27.3 betas, with App Intents and new multi-step "skills"
Apple's Developer Preview lets any app expose its content and actions to Siri AI through App Intents, so people can ask in plain language and Siri runs your custom action. New skills let you chain several actions into one request, and Siri AI can now surface any entity in the Spotlight semantic index. It ships to customers in a few months and App Store submission opens in the spring. Supported languages are English, French, Japanese, Korean, Portuguese and Spanish, and it isn't available in the EU on iOS yet.
Link: developer.apple.com/news
Try it now
- Install the iOS 27.3 beta on a test device and read the Siri AI Developer Preview docs.
- Ask Siri AI "what can you do in Notes?" to see Apple's own reference actions.
- Point your coding agent at the project: "List every App Intent this app already defines, then propose the three user requests in plain English that should map to them, and one skill that chains two of them."
- Apple's Group Labs run November 4 and 5 if you want to ask an engineer.
Story II
Anthropic open-sources FlashProve, which proves a reply came from a committed model at about 110 tokens/second on one H100
FlashProve generates zkSNARK proofs that a specific set of weights produced a given output, without revealing those weights. On Llama 4 Maverick, a 400B mixture-of-experts model in 4-bit, one H100 with 24 CPU cores proves about 110 tokens/second. Each 1,024-token request yields an 18.78 MB proof that verifies on a CPU in 17 seconds. The catches are real: contexts top out at 1,024 tokens, decoding is greedy, and proving costs roughly 70 to 95 times the GPU time of plain vLLM. It's Apache 2.0, published as a research artifact that isn't taking contributions.
Link: github.com/anthropics/FlashProve
Try it now
git clone https://github.com/anthropics/FlashProve && cd FlashProve
Start with guide/performance.md for the batch-size numbers, then the README's verifier section. You can run the verifier on a laptop CPU. Proving needs an 80 GB H100 and about 290 GB of host RAM, so price that out before you plan a demo.
Story III
Soniox STT v6 reads what your voice agent is saying, then uses it to hear short replies and filter the agent's own echo
With "enable_agent_text_input": true on stt-rt-v6, you stream the agent's words over the same WebSocket as the caller audio. The model uses that text as context, so "River Roast" stops turning into "River Road", and it filters leftover agent voice that your acoustic echo cancellation misses. It covers 60+ languages at about $0.12 an hour streaming and $0.10 async. Soniox hasn't published v6 accuracy numbers yet, per MarkTechPost. The v5 models go away on December 15.
Link: soniox.com/blog
Try it now
- Change the model name:
{"model": "stt-rt-v6"} or {"model": "stt-async-v6"}. - Add
"enable_agent_text_input": true to the start request. - Send each TTS chunk as
{"type": "agent_text", "text": "...", "end_of_segment": true} 500 ms to 1 second before playback. - Test money formatting first: v6 writes "15 dollars and 20 cents" as spoken.
Story IV
Tencent's WeVisOne-VL does detection, OCR, counting, segmentation and alpha matting in one Apache 2.0 model
WeVisOne-VL comes in 2B and 4B sizes, built on Qwen3-VL, with a 9.6M-parameter pixel decoder that replaces a separate SAM for masks and mattes. When the object you asked for isn't in the image, it answers None instead of inventing a box. Tencent reports 71.3 F1 on Dense200 for the 4B model, against 58.7 for LocateAnything-3B. These are self-reported numbers. The repo has about 875 stars and hosted demos for both sizes.
Link: github.com/Tencent/WeVisOne-VL
Try it now
No GPU handy? Try the 4B demo on Hugging Face. Locally:
git clone https://github.com/Tencent/WeVisOne-VL.git && cd WeVisOne-VL
pip install -r requirements.txt && pip install -v -e .
Then in Python: WeVisOneVL("tencent/WeVisOne-VL-2B", backend="transformers").predict(image, task="matting", query="person")
— The Vibe Gate news desk. We read the firehose so you can keep building.