Matching stories Newest first
Hugging Face model Spark-X2.6-4B, a compact on-device language model, is released
Spark-X2.6-4B keeps the Spark-X2.5-4B architecture and gains from improved post-training, with three thinking levels (none, low, high). The developer claims it matches or beats models around 9B parameters on open-source benchmarks.
Gradium releases Phonon, open-source on-device text-to-speech models
Gradium released Phonon, two small TTS models of 31M and 40M parameters, with a fully open-source inference stack targeting CPUs, GPUs and NPUs across 10+ backends, mostly written in Rust.
Pocket TTS with voice cloning shrinks to 1.9 million parameters
A community post says the Pocket TTS text-to-speech model, which supports voice cloning, is now only 1.9 million parameters.
Six-way benchmark compares Qwen3.8-Flash-Next locally against Claude Opus 5.5 on a Rust feature
A community test of 12 PRs on a LlamaStash feature found Opus 5.5 averaged 82% against 68% for Claude Code with Flash-Next and 66% for Pi with Flash-Next. Flash-Next ran locally at about 4.16 bpw, so the author paid only electricity, and results are a single developer's unverified experiment.
Ornith-1.5-122B, a Qwen-based finetune, released in GGUF and NVFP4 formats
Ornith AI introduces Ornith-1.5, which continuously generates its own training tasks and improves through reinforcement learning, extending Ornith-1.0 built on Qwen3.5 and Gemma4. GGUF and NVFP4 builds are on Hugging Face, with more details in the company blog.
ISTA-DASLab releases a Qwen3.8-27B prefiller model that streams from SSD for low-quant accuracy
The demo recovers accuracy for low-bit quantized Qwen3.8-27B while keeping VRAM usage low and decoding fast. It reportedly needs an Nvidia Blackwell GPU, and the poster has not tested it. For context longer than about 4k tokens, the poster expects no speed penalty.
HermesConsole and VoiceManager open-source a low latency local speech-to-speech demo
A hobbyist released VoiceManager, a streaming STT and TTS server under MIT, with a Python client and a Kotlin console reference app. Barge-in is off by default because echo cancellation is hard, and the default setup uses Nemotron and Breeze TTS 2.
Tess-4.5-27B, a Qwen finetune, claims 25% fewer tokens at the same score
Community finetune by migtissera built on Qwen3.8-27B, with GGUF files on Hugging Face. The author reports 25% fewer tokens across 150 BenchLocal scenarios and 43% fewer on data extraction, with the same scores claimed.