ERNIE 5.5 Flash Base, Phonon TTS, CUDA 13.5, Luma MCP

Four primary pages I could open and read today: an open base model, a small speech model, a toolkit release and a connector. Everything below comes from those pages, not from my own testing.

ERNIE 5.5 Flash Base from Baidu

The Hugging Face card describes ERNIE 5.5 Flash as a sparse mixture-of-experts language model with 280B total parameters and 14B activated per token. It was pre-trained from scratch on 33.5T tokens with a training context of up to 128K, and fine-tuning and inference support 1M tokens.

What I find useful is the choice to release the checkpoint at the end of pre-training, before any mid- or post-training. Baidu says it is meant as a common starting point for domain adaptation, instruction tuning and reinforcement-learning research. The pre-training data leans toward repository-scale code and long-context understanding, and the card lists efficiency, meaning activated compute and KV cache cost, as a design goal from the start. Baidu compares it against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base under one evaluation pipeline. FP8 and BF16 safetensors are both offered, and the license field reads "other", so I would read the terms before building on it.

Phonon TTS from Gradium

Gradium's post is a look back at a launch two weeks ago. Phonon comes in two sizes, about 65 MB and 75 MB, and runs on phones, laptops, single-board computers and in a browser tab with no server. Phonon 31m speaks English under Apache 2.0. Phonon 40m covers English, French, German, Spanish and Portuguese under a non-commercial license.

On the Hugging Face Open TTS Leaderboard, Gradium says Phonon ranks first in German and Spanish for intelligibility and in the top three in every language it evaluated, next to models up to 100 times its size. It also tops the CPU streaming benchmark on time to first audio. The team reports more than 200,000 downloads in the first two weeks, a new inference framework called ptts with more than 10 backends, and a reference implementation in Transformers from v5.21.

CUDA 13.5 from NVIDIA

Vera Rubin support moves from early developer access to general availability, and Windows on Arm support now extends to DGX Station for Windows. The scheduling changes are the part I would test first. A new flag lets locality-domain partitions include the supplemental SMs, so an application can trade uniform partition sizes for using every SM on the device. MPS v3 adds weighted timeslicing, so servers that need longer compute turns can get them.

The libraries move too. CUDA Tile gains multi-GPU communication over NVLink and NVSwitch, and CCCL 3.5 adds batched Top-K selection integrated into FlashInfer and reproducible floating-point scans.

Luma MCP connector

Luma now runs inside Claude and ChatGPT through an MCP connector. The pitch is one winning ad becoming a dozen variants without leaving the chat: product shots, new camera angles, background swaps, relighting and short promo clips. It uses your Luma credits, so there is no separate billing.

What I'd watch next

I want to see whether anyone publishes instruction-tuned or reinforcement-learning results on top of the ERNIE base, and whether Phonon's leaderboard standing holds as more open TTS models are added.