Daily bookmarks: relevant news, facts and discoveries.
Articles 1219
Engineering ·
A post claims Meta's paper 'Breaking the Token Ceiling' trains models on raw bytes with a 256-value vocabulary and applies distillation. According to the post, 1B byte models match or beat two cited 1B models. No link to the original work was provided.
Engineering ·
An author claims that TLS inspection roots installed on corporate laptops can issue certificates for any site without appearing in certificate transparency logs, since browsers require that only of publicly trusted CAs. The source gives no links.
Engineering ·
A Swift port of the open Phonon 31M text-to-speech model, running CoreML on Apple's Neural Engine, with 103 voices and a text normalizer ported from Gradium. The author reports about 20 ms to first audio and about 20x real time on an M5 Pro.
Engineering ·
A post from the Tongyi Speech AI account explains streaming speech recognition, which must define words while the speaker is still talking, and how latency and lookahead shape model and decoder design.
Engineering ·
Kyutai Labs says it released smaller Pocket TTS versions at 18.2M, 6.9M and 1.9M parameters, with voice cloning and CPU execution. The source reports 103x real-time speed for the 1.9M model on a single core.
Engineering ·
A large language model lab released as open source the inference framework it uses to serve its models in production, described as high-performance and large-scale.
Engineering ·
Open-source iOS project in which an angel and a devil on your shoulder answer questions by voice, with speech-to-text and text-to-speech running on the phone itself, without network inference, according to the source dated October 20, 2026.
Engineering ·
A post describes diffusion text models, which produce the response simultaneously instead of token by token as autoregressive LLMs do. The author says generation is fast, but the text gives no measurements. Published date: 2026-10-20.
Engineering ·
An author announces open source text-to-speech models about half the size of Kokoro (82M parameters), reporting higher scores for naturalness and accuracy and support for voice cloning. No link is provided, so the results have not been verified.
Engineering ·
Radar: the author announces Phonon, a text-to-speech model of about 40M parameters, multilingual, with a voice library and local use through the speech-to-speech package.
Engineering ·
A post reports porting Phonon, an open text-to-speech model with 31M parameters from GradiumAI, to run fully in the browser via WebGPU in half precision, at about 47x real time on an M4 Pro and 7x on a Phone 13 Pro.
Engineering ·
Post on Intel's D270 (Crescent Island) GPU, a PCIe 5.0 card for inference with 32 Xe cores, XMX matrix engines and about 160 GB of LPDDR5X expandable to 480 GB, and its possible effect on large models and cost per token.
Engineering ·
Ant Ling announced the open-source release of five Ming-3.0 speech models, including TTS variants, audio and dialogue understanding models, and an ASR model for dialects. The source provides no links.
Engineering ·
An English speech model with 949 thousand parameters for CPU, reported by the author at about 436 times real time and 11.43% WER on test-clean.
Engineering ·
A pull request adds the Drama-3 model, from Fish Audio, as a streaming speech synthesis option in a speech-to-speech pipeline, with optional voice cloning and expressive intonation.
Engineering ·
A reviewer reported 18 to 58 tokens per second with four large local models on llama.cpp with Vulkan, then returned the hardware because of its cost.
Engineering ·
Report of an AI agent using the open OpenVoice model, from MIT and MyShell, to dub a Polish film with synchronized dialogue on a 16GB Mac mini M4.
Engineering ·
Open source repository letting AI agents control real Chrome sessions via MCP and the Chrome DevTools Protocol, with navigation, network hooks and DOM extraction. The author claims it bypasses Cloudflare challenges.
Engineering ·
On October 15, 2026, IBM published Granite Switch 4.1, in 8B and 30B, combining chat, decision and safety adapters and speech recognition with about the same GPU memory as the base model.
Engineering ·
On October 16, 2026, Intel released version 1.5 of LLM-Scaler-vLLM, a Docker stack with vLLM for Intel Arc GPUs, which jumped from vLLM 0.14 to 0.26 and added more models and FP8 quantization.
Looking for something else?DocumentationContact
