MTP nearly doubled Qwen 3.8 27B on my 4090.
It also filled the card to the last 112 MB. Qwen 3.8 dropped today with a draft layer baked into the weights. The idea is simple: a small attached head guesses the next couple of tokens, the main model checks all the guesses in one
Qwen 3.8:27b (via Ollama) It works fine, but not at a practical speed...lol
LFG! one shot. updated harness + qwen 3.8 + minimax-music3 + ltx 2.5.
Full music video. not perfect. not bad.
good morning. This is the AI morning edition of August 15th. ① Alibaba “Qwen 3.8” 27 billion parameters released with Apache 2.0
②GPT-5.6 Sol Ultrafast mode, up to 750 tokens/sec with Cerebras partnership ③ Google turns off visible watermark on AI-generated content You can catch up on the influence and background in just 4 minutes of video. (YouTube link)
You have reached the end of the archive
All of qwen38