Sped up Qwen3-ASR 1.7B to beat deepgram (www.reddit.com via reddit)
model roundup
Qwen 3
-
Guys I solved a bunch of problems and was able to improve the Latency of Qwen to match that of deepgram, for streaming voice calls. Since Qwen is an more of an audio LLM ASR model, it has the smartness that ASR needs.
-
Does it Exist: A Leaderboard for Speech Recognition / STT / Transcribing? (www.reddit.com via reddit)
My last post said I'm integrating simplex with dsh, and I'm working on a voice to text feature for it. Dsh implement it with a fallback if local whispercpp fails.
-
Running Qwen 3 27B on 3090 or Mac or whatever? (www.reddit.com via reddit)
SOO im seeing soo much hype on this model and im seeing everyone be running it on anything, im very curious. I would like to try to run in but in all honestly i know there's like soo many quantized version...
-
Teaching a local LLM to reason about a new domain through continued pretraining (www.teachmecoolstuff.com via hn)
<p>As an interesting experiment I wanted to learn how to teach a tiny local llm a new domain by doing Continued Pretraining (CPT) on domain specific data. This article is a write-up of my experiences from using Unsloth to train qwen 3 4B t…
-
Show HN: Ollama served my 40k-context model at 4k, silently (github.com via hn)
A model that advertises 40,960 tokens was being served 4,096 I've been running knowledge distillation experiments on consumer hardware: one RTX 5080 (16GB VRAM), 32GB of system RAM, an 850W supply, and Qwen3 at 1.7B, 4B, 8B, and 14B, all Q…