AI Briefing#qwen3
The latest AI and developer news about #qwen3, with the original source and a short summary.
The latest AI and developer news about #qwen3, with the original source and a short summary.
The Apple Silicon-specific inference engine Lily achieved a 35% improvement in decoding speed and high hardware utilization compared to MLX-LM.
The Qwen3-Coder-Next 4bit model, quantized using the calibration-free TQ method, has been released, reducing disk usage to 41.7GB.
vLLM resolved the non-deterministic output issue for Qwen3.8-Flash-Next at temperature 0 by fixing the sparse-attention kernel.
Optimized Qwen3-TTS to achieve sub-50ms latency and 10 RPS throughput on a single H100.
Unsloth has released Dynamic v3.0 quantized models for Qwen3.8-27B, demonstrating improved accuracy.
Technical specifications have been released for Neutrino-1 8B, a dense decoder-only transformer model built on Qwen3-8B.
llama.cpp added MTP support, and Qwen3.6 MTP GGUF models were released.
Orthrus, based on Qwen3, achieved up to 7.8x speedup through lossless parallel generation.
Orthrus claims up to a 7.8x speedup on Qwen3-8B inference while keeping the backbone frozen.
AutoTTS discovers Test-Time Scaling policies through code exploration alone in an offline replay environment.
A llama.cpp regression increased the Qwen3.6-27B IQ4_XS size from 14.7GB to 15.1GB.
Compared the performance of Qwen 3.6 27B's BF16, Q4_K_M, and Q8_0 quantizations.
On 4x Tesla M40 12GB with an fp4 setup, Qwen3.6 27B speed was reported at 26 tok/s PP and 8 tok/s TG.
Adversarial user prompts consistently lowered IFEval performance across 14 LLMs.
Running Qwen3.6-27B on a single RTX 3090 achieved 85 TPS at 125K context.
In an apples-to-apples FP8 comparison, the advantage between Ironwood and B200 reversed depending on the model.
In Open WebUI with OpenUI attached, gpt-5.4-mini responded the fastest.
ElevenLabs has introduced self-hosted LLMs into its agent platform, reducing latency and improving cost efficiency for voice agents.
Furiosa SDK 2026.1 strengthens batching, caching, and Kubernetes support.
Based on Qwen3-4B, DeepMath uses Python code to improve math reasoning accuracy while cutting output length by 66%.
Jamba Reasoning 3B was much faster than Qwen3 4B on a 60K-token QA task.
OpenVINO and depth-pruning boosted Qwen3-8B inference speed by 1.4x on Intel Core Ultra.
Qwen3Guard is a safety guardrail that judges prompts and responses in real time.
GSPO reduces GRPO's instability through sequence-level optimization, enabling RL scaling.