MOSS-Transcribe-Diarize 0.9B — An End-to-End Speech Transcription Model Supporting 50+ Languages
Ranked among the top trending models on Hugging Face, MOSS-Transcribe-Diarize 0.9B is an end-to-end model that transcribes long, multi-speaker audio in a single pass. It was developed by OpenMOSS-Team and released on July 9, 2026.
Its biggest strength is that it integrates ASR (automatic speech recognition) and diarization (speaker separation), which were previously performed separately. For long recordings such as meetings, podcasts, interviews, and lectures, it outputs a structured transcript complete with timestamps and speaker labels ([S01], [S02], etc.) in one pass. It supports more than 50 languages.
The model architecture combines a Qwen3-0.6B-based text backbone with a Whisper-Medium encoder, packing everything into a compact 0.9B size. In benchmarks, it records accuracy surpassing Doubao, Gemini 3 Pro, and GPT-4o on multi-speaker datasets such as AISHELL-4 and Alimeeting.
It is also well thought out for practical use: in addition to local Python inference, it can be served as an OpenAI-compatible API through vLLM or SGLang Omni, and a web app for subtitle creation (mtd-subtitle-web) is also provided. It is released under the Apache License 2.0, so commercial use is possible.
AI Agent Beats Slay the Spire 2 with Structured Memory
The AgenticSTS project, reported by The Decoder, presents a new approach to the context-length problem that AI agents face.
Most AI agents accumulate their action history as a chat log, so the prompt keeps ballooning over long tasks. AgenticSTS replaced this chat log with five independent memory layers. As a result, it kept the prompt size down to about 5,000 tokens while achieving 6 wins out of 10 matches in the card game Slay the Spire 2. The existing agents it was compared against could not win a single game.
This result demonstrates that an agent's memory design directly affects task performance, and it could also influence practical applications that require long-horizon task execution — not just games.
Xiaomi Quietly Releases MiMo-V2.5-DFlash Weights
The LocalLLaMA community on Reddit discovered that Xiaomi has published the weights of MiMo-V2.5-DFlash on Hugging Face. MiMo-V2.5 is a large-scale model with over 300B parameters, and it is reportedly running at 8-10 tokens per second on a 2x24GB GPU + VRAM offload setup.
The application of DFlash technology is expected to improve inference speed. A model with an MTP (Multi-Token Prediction) head has also been released separately, and the community is eagerly awaiting an implementation in llama.cpp.
Voodoo Quant — A New Quantization Method That Improves on Unsloth Dynamic 2.0 by 95%
Also announced in the LocalLLaMA community, Voodoo Quant shares the same direction as Unsloth Dynamic 2.0 (allocating higher precision to important tensors), but differs in that it optimizes every single tensor individually rather than working at the tensor-group level.
In tests on Qwen3.5 0.8B and 2B, it improved KL divergence (the difference between distributions) by 95% compared to Unsloth Dynamic 2.0. Particularly interesting is that while Unsloth Dynamic performs well on llama.cpp but degrades on the PyTorch graph, Voodoo Quant works in a balanced manner across both environments.
2-bit precision is reported to be the "sweet spot," and promising results are expected when it is applied to practical models such as Qwen3.6 27B and DeepSeek V4 Flash in the future.
---