Qwen announced Qwen3.8-LiveTranslate, a real-time simultaneous interpretation model built on an Interleave architecture. The model improves faithfulness, fluency, and conciseness while cutting average lagging (LAAL) from 2.8s to 2.3s across 60 languages. For practitioners, it targets lower-latency live translation across a broad multilingual footprint.
Qwen announced Qwen3.8-Omni-Flash, its first omni-modal model designed around agentic capabilities. The model combines native audio-video understanding, reasoning, and tool use so it can understand content, plan a task, execute with tools, and deliver a result. For practitioners, it signals Qwen's move toward single-model agents spanning perception and action.
Qwen announced Qwen3.8-Omni-Flash, its first omni-modal model designed around agentic capabilities. The model combines native audio-video understanding, reasoning, and tool use so it can understand content, plan a task, execute with tools, and deliver a result. For practitioners, it signals Qwen's move toward single-model agents spanning perception and action.
A new arxiv paper proposes a data-driven method for simultaneous speech translation using decoder-only LLMs, avoiding architectural changes. The approach uses fixed-length chunks, cumulative streaming decoding with a rewind-based committed prefix, and teacher-labeled prefix-to-prefix (P2P) training to handle cross-lingual reordering under latency constraints.
A new arxiv paper proposes a data-driven method for simultaneous speech translation using decoder-only LLMs, avoiding architectural changes. The approach uses fixed-length chunks, cumulative streaming decoding with a rewind-based committed prefix, and teacher-labeled prefix-to-prefix (P2P) training to handle cross-lingual reordering under latency constraints.
Qwen3.8-27B, a 27B-parameter open-weight model, reached #1 on Harvey's Legal Agent benchmark and became the top trending model on Hugging Face. It also hit #1 on Cline within four days, signaling strong community adoption for a local model.
Qwen3.8-27B, a 27B-parameter open-weight model, reached #1 on Harvey's Legal Agent benchmark and became the top trending model on Hugging Face. It also hit #1 on Cline within four days, signaling strong community adoption for a local model.
A user translating 670k short English product descriptions into five languages finds Qwen3.6 27B and 35B too slow at 60–70 TPS, estimating 40 days for the full set. They are exploring smaller quantizations, MTP, and alternative models to speed up the batch translation task.
A user translating 670k short English product descriptions into five languages finds Qwen3.6 27B and 35B too slow at 60–70 TPS, estimating 40 days for the full set. They are exploring smaller quantizations, MTP, and alternative models to speed up the batch translation task.