Ali Asif

The AI log, open weights · agents and coding · image, video, world

ZONOS3-Mini, BotPsy, NVIDIA Snapshot, Seedream 5.5

An open 100M-active speech model, a 0.8B tool-use model, GPU checkpoint and restore for Kubernetes, and Seedream 5.5 in ComfyUI.

By , published

What shipped

  1. Zyphra

    ZONOS3-Mini

    An Apache 2.0 text-to-speech model with 100 million active parameters, 43 languages and voice cloning, running on phones, laptops and in the browser via WebGPU.

  2. Tether AI Research (QVAC)

    BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

    A 0.8B model built on Qwen3.5-0.8B that roughly doubles its base on tool-use benchmarks, for example BFCL v4 from 23.4 to 44.0.

  3. NVIDIA

    Start and Restore GPU Workloads in Seconds with NVIDIA Snapshot

    An open-source checkpoint and restore system for GPU workloads on Kubernetes that brings inference replicas online in 3.6 to 42.7 seconds, 8x to 36x faster than cold starts.

  4. ComfyUI

    Seedream 5.5 is now available via Partner Nodes

    Seedream 5.5 arrives in ComfyUI with multi-turn localized editing, native transparent RGBA output and readable in-image text.

Four primary pages I opened today, all about making models smaller, faster to start or easier to wire into a workflow. Everything below comes from those pages, not from my own testing.

ZONOS3-Mini

Zyphra describes ZONOS3-Mini as a compact, open-source text-to-speech model with 100 million active parameters. The weights are under Apache 2.0. It supports 43 languages, clones a voice from a short reference clip, and lets you shape age, accent and vocal texture for a designed character. Emotion, pace and delivery can be steered down to individual words.

The deployment story is what stands out. The card lists a WebGPU build that runs entirely in the browser with no server, a CUDA server, an MLX server for Apple Silicon, and a C++ ggml server that covers CPU, Metal and Vulkan. SGLang-Omni can also serve it behind an OpenAI-compatible speech endpoint with streaming. A hosted version exists on Zyphra Cloud for people who prefer that.

BotPsy

Tether AI Research introduces BotPsy, a 0.8B model for function calling and agentic work, built on Qwen3.5-0.8B and aimed at on-device use. On the team's evaluations it roughly doubles its base: BFCL v4 goes from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and tau-bench from 23.2 to 51.5. It leads five of six benchmarks even against Qwen3.5-2B-Think, a model 2.5 times its size, which still wins on Claw-Eval.

The authors are plain about limits. Multi-step information retrieval is the weak spot, performance drops as instructions pile up, and lower-resource languages lag. They say it is not meant for high-stakes use. GGUF files in bf16, Q8 and Q4 are linked.

NVIDIA Snapshot

NVIDIA Snapshot is a new open-source component of its DSX AI Factory Platform. It captures initialized CPU and GPU memory state for a workload on Kubernetes, so a later replica restores from that artifact instead of cold starting. Across tested configurations, restores bring inference replicas online in 3.6 to 42.7 seconds, an 8x to 36x speedup over cold starts that take minutes.

It handles multi-GPU setups, including tensor, data, expert and decode context parallelism, with checkpoints up to 1 TB across eight GPUs. Checkpoint creation takes 15 seconds to 5.5 minutes, and checkpoint plus restore beats a single cold start in every tested case. These are NVIDIA's numbers on B200 GPUs with VAST NFS storage.

Seedream 5.5 in ComfyUI

ComfyUI added Seedream 5.5 as a Partner Node, so it runs behind one node and uses credits. The post lists multi-turn editing that changes only the region you describe, native transparent and semi-transparent RGBA output, posters and infographics with readable text, and lighting control from the prompt. No benchmarks are given, so this is a feature list rather than a measured result.

What I'd watch next

I want independent numbers for BotPsy on tasks outside its own benchmark set, and a look at how ZONOS3-Mini sounds in languages beyond English. For Snapshot, I would watch whether the restore times hold on storage other than the one NVIDIA tested.

Sources

Topics: NVIDIA, Kubernetes, ComfyUI, Inference, ON Device, Image Generation, Text TO Speech