AI Briefing#qwen
The latest AI and developer news about #qwen, with the original source and a short summary.
The latest AI and developer news about #qwen, with the original source and a short summary.
Using turbo3, achieved about 40 tok/s running Qwen3.6 35B A3B at 260k ctx on a 12GB VRAM 3080.
Qwen3.6 adds preserve_thinking, improving reasoning context and KV cache efficiency.
Swapped Qwen 1.5B's decoder head with a CHIP-8 opcode head and ran it directly.
Qwen3.6-35B-A3B produced a better pelican SVG than Claude Opus 4.7 on a local laptop.
On refined data, 1.7B beat 744B, but noisy traces collapsed performance.
Qwen3.6 has been released with enhanced frontend workflows and repo-level reasoning.
Qwen released a MoE model with 35B total parameters and 3B active parameters under Apache 2.0.
Gemma 4 31B produced better results than competing models with fewer tokens in an image-based 3D geometry test.
Trained Qwen2.5-0.5B-Instruct with GRPO on three Mac minis and compared summary quality.
Compared MMLU scores of GGUF quantized models in a 24GB VRAM environment.
On an RTX 5080·64GB Windows 11 setup, Llama.cpp was about 2x faster than LM Studio.
Results from training a shopping agent with RL across 8 verifiable e-commerce environments.
A summary of how to directly train multimodal embedding and reranker models using Sentence Transformers.
Qwen3-1.7B surpassed GLM-5 744B using only synthetic data fine-tuning.
A tuning log running Qwen3.5-35B at 64k context on an RTX 4060 Ti 16GB, hitting 40-60 tok/s.
Gemma4 26B and E4B reportedly replaced Qwen 3.5's routing/response quality issues in practice.
Built a paper-based web app with Qwen3.5 35B and confirmed it maintains context well up to 70-80K tokens.
Dropping stateless execution in favor of attached sessions stabilized 3-agent collaboration.
In the latest oMLX commit, DFlash optimization significantly boosted tg TPS for small contexts.
On an M2 Max, DTree showed MLX results **1.07x** faster than DFlash.
Built a memory RAG that lets you query all your personal data using Ollama and qwen2.5:7b.
Shares how to split up models by role for OpenClaw based on PinchBench scores.
Trained Qwen2.5-0.5B-Instruct on a Reddit summarization task using GRPO.