NVIDIA's Indic corpus, MOSS-Image, Hy-Memory and Magnetar

No single frontier launch today. What I found is a set of smaller releases that each come with something concrete to read: a recipe, a model card, a paper-backed claim. I opened all five pages before writing this.

NVIDIA's Indic pretraining corpus and the recipe behind it

The biggest item is NVIDIA's post on Nemotron-pretraining-Indic-v1. It describes about 2.8 trillion tokens across 22 Indian languages, built with IIT Madras and its AI4Bharat lab. The post says the corpus is "releasing soon", so today the recipe is the part you can use. The data is 15 Nemotron-CC-v2 subsets translated from English, with roughly 2.6 trillion native-script tokens and 224.6 billion romanized ones.

For twelve languages, Gemma-4-31B-it translates whole documents and then scores each translation against its source with a rubric. Anything under 75 out of 100 is repaired or dropped, and the repair step recovered 438.4 billion tokens. The other ten languages use a dedicated translation model, with Gemma-4-31B-it as the judge. The prompts, quality checks, windowing and script remapping are published as a NeMo Data Designer plugin on GitHub.

On results, the post says Nemotron 3.5 Super VL gained between 4.98 and 9.82 points across five Indic benchmarks after continued pretraining. It is careful to say that this mixes synthetic and natural data, so the corpus alone does not explain the gain. It also lists limits: human review varies by language, and translated text does not replace text people wrote.

MOSS-Image: a 5B open image model with layout control

OpenMOSS released MOSS-Image-SFT, fine-tuned from MOSS-Image-Base. The card lists about 5B parameters and an Apache 2.0 license. You can prompt it in plain language, with tag sequences, or with a JSON layout that fixes where text and objects go. It supports English and Chinese, has a Diffusers pipeline, and a ComfyUI plugin you install through ComfyUI Manager. The card does not say what data it was trained on, so I would not assume anything there.

Hy-Memory: a small model for deciding what to remember

Tencent's Hy-Memory-30B-A3B is a memory writer. It sits inside frameworks such as Mem0, MemOS and Hy-Memory itself and decides what to keep, how to store it and how new evidence updates the bank. The card describes two training stages, on-policy distillation and then reinforcement learning, and says it transfers to EverMemOS, which was held out. The benchmark numbers are self-reported on fixed 20 percent subsamples, with an overall average of 55.49 against 54.62 for Claude Opus 5. I would run it on my own memory traces before believing the comparison.

Magnetar and Instant Soundtracks

Microsoft Research's Magnetar is a zero-knowledge VM that proves RISC-V programs, so Rust, Go or C compile as usual. The post reports 7.2 to 8.7 million instructions proved per second on a laptop, proofs that verify in about 12 milliseconds, and security that rests only on hash functions. The post does not say whether the code is open.

ElevenLabs shipped Instant Soundtracks in its Music Video Editor and the Studio API. It scores an uploaded video, adds sound effects at matching timestamps, and can return music and effects as separate tracks. The limits are 10 videos, 200 MB and 10 minutes per request.

What I'd watch next

Whether the Indic corpus actually lands on Hugging Face, and whether the plugin works on a language I can check. Then whether Hy-Memory holds up outside the lab's own test setups, and whether Magnetar gets a public repo.