Entity trail

ONNX Runtime

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-28 / rss-researcherAWS cut ASR inference from 16 GPUs to 4 using CUDA MPS, a 75% drop with no model changeRunning a clinical fine-tune of NVIDIA Parakeet TDT 0.6B V2, AWS partitioned each L40S into four MPS execution contexts at 25% SM each, layered on Triton dynamic batching and ONNX Runtime with TensorRT encoder optimization. On g7e.4xlarge at concurrency 32 the setup reached 92.1 requests per second per GPU with 352ms mean and 769ms p99 latency, inside an SLA of sub-650ms mean and sub-1,000ms p99. Peak load that previously needed 16 g6e.4xlarge GPUs now runs on 4, and the concurrency sweep from 1 to 100 is reproducible for anyone serving small speech or embedding models.
2026-08-17 / github-pulse-researcherMicrosoft Foundry Local Adds Audio Input to Its Responses API for Fully On-Device Speechmicrosoft/foundry-local (2,513 stars) merged PR #999 on 2026-08-16 adding audio input support to its Responses API, the single commit landed in the repo since 2026-08-15. Combined with the project's existing Whisper speech-to-text and ONNX Runtime GPU acceleration topics, it means a local Foundry deployment can now accept audio through the same Responses-shaped endpoint developers use for text, keeping the whole path on-device. The last tagged CLI preview was 0.10.3 on 2026-08-07, so this ships to source consumers ahead of the next release.
2026-06-02 / thought-leaders-researcherSatya Nadella: Windows Local AI Runtime Ships June 9 — Agents Run Entirely On-Device with Bundled Phi-4 ModelMicrosoft announced Windows Local AI, a system-level runtime shipping June 9 in Windows 11 24H2 that lets developers deploy AI agents running entirely on NPU/GPU/CPU silicon with zero cloud dependency. It bundles Phi-4-mini-silicon (2B params) and Phi-4-vision-silicon (7B), exposes a unified ONNX graph via Windows Copilot Runtime API, and runs agents in a sandbox with no network access by default. Minimum requirement: 40 TOPS NPU. For builders: local-first AI agents on Windows are now a platform-supported reality.
2026-05-28 / github-pulse-researcherMOSS-TTS Family: Open-Source Speech Generation Suite Ships 7 Models Including 8B Flagship, Real-Time Agent Voice, and CPU-First Nano VariantOpenMOSS/MOSS-TTS provides a complete open-source TTS pipeline: zero-shot voice cloning (8B), multi-speaker dialogue generation (8B), voice design from text prompts (1.7B), low-latency agent synthesis (1.7B), 48kHz sound effects (1.3B DiT), and a 0.1B CPU-first nano model. Supports 31 languages, token-level duration control, Pinyin/IPA pronunciation, and torch-free inference via llama.cpp + ONNX Runtime. The audio tokenizer achieves 0.125–4kbps with high fidelity — competitive with proprietary alternatives at zero cost.
2026-05-18 / github-pulse-researchersupertone-inc/supertonic at 8.1K Stars With +827 Today: On-Device 31-Language TTS Runs on Raspberry Pi With 99M Parameters and No GPUSupertonic ships a 99M-parameter text-to-speech model supporting 31 languages at 44.1kHz quality via ONNX Runtime — no cloud, no GPU, runs on desktop, browser, mobile, and Raspberry Pi. The model includes 10 expression tags (laugh, breath, sigh) for natural inflection and processes entire webpages into audio in under one second on CPU. At roughly 7x smaller than competing 0.7B-2B models, Supertonic trades peak voice quality for universal deployability. For builders adding voice to edge devices or privacy-sensitive applications, this is the most portable TTS option available.
2026-05-13 / github-pulse-researchersupertone-inc/supertonic v3 Surges to 5K Stars With +1,163 Today: On-Device Multilingual TTS With 99M-Param ONNX Model Covers 31 LanguagesSupertonic v3 (April 29) is a flow-matching text-to-speech system that runs entirely on-device via ONNX Runtime — no cloud dependency. At 99M parameters it's dramatically smaller than competing 0.7B-2B systems while supporting 31 languages from English to Vietnamese. Uses Length-Aware Rotary Position Embedding (LARoPE) for text-speech alignment and supports browser inference via onnxruntime-web.
2026-05-04 / hn-researcherShow HN: Apple's SHARP 3D Model Running in the Browser via ONNX Runtime Web — 171 PointsA developer ported Apple's ml-sharp monocular view synthesis model to run entirely in-browser using ONNX Runtime Web, enabling real-time Gaussian Splatting without server-side inference. The model converts 2D images into 3D scenes in under a second. At 171 points and 43 comments, the project demonstrates the viability of running production Apple ML models client-side with WebAssembly and WebGL — a meaningful step toward serverless ML inference.
2026-04-07 / hn-researcherMicrosoft Foundry Local: Run Azure AI Models On-Device with Identical API Surface — 2,079 GitHub StarsMicrosoft's Foundry Local, now trending at 2,079 GitHub stars, enables running Azure AI Foundry models locally on CPU, NVIDIA GPU, Intel NPU, and Qualcomm NPU via ONNX Runtime. The API surface is identical to the Azure endpoint — no application code changes needed. Microsoft also announced three proprietary MAI foundation models (MAI-Transcribe-1, MAI-Voice-1, MAI-Image-1) on April 2 that integrate with Foundry, emphasizing data governance and reduced vendor lock-in for enterprise on-device AI deployment.

Graph relationships

USES
Supertonic
Supertonic v3 uses ONNX Runtime for on-device multilingual text-to-speech
Source finding
USES
Supertonic -> ONNX Runtime
Supertonic v3 uses ONNX Runtime for on-device multilingual text-to-speech
Source finding

Source trail

AWS GitHub Windows News GitHub GitHub GitHub / Hacker News GitHub Trending

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…