Tag
All content about Qwen, organized for fast scanning.
Qwen has introduced its latest multimodal model, Qwen3.8-27B, featuring a 262K-token context window and enhanced performance capabilities. Additionally, Alibaba launched Qwen3.8-Max, a significantly larger 2.4 trillion-parameter model designed for complex tasks, while Qwen is engaging developers through a feedback program that offers token rewards. The company also launched a new subscription plan for Qwen Cloud, which consolidates access to various models and has sparked discussions about pricing and service features.
DFlash 2 claims it can run Qwen3.8-27B at up to 70 tokens per second on an M5 Max MacBook Pro, with “up to 4.6×” faster decoding at the same output. Early user reports vary widely, and cross-hardware comparisons remain unclear.
Qwen has just rolled out Qwen3.8-27B, an Apache 2.0-licensed 27B multimodal model with a native 262K-token context window. The company says it beats Qwen3.7-Plus, and early testers are already pushing local builds and quantized runs.
Alibaba has just rolled out Qwen3.8-Max, a 2.4 trillion-parameter model aimed at coding, workplace tasks, and multimodal agent workflows. Qwen claims it can run autonomously for days on long projects, with open weights promised next week alongside Qwen3.8-27B.
With the launch of Qwen3.8-Max-Preview, Qwen is inviting developers to submit real-world “good or bad” agent task cases by August 7. The #QwenGrowthPlan offers token rewards and potential promo placement as users ask for smaller variants, fewer limits, and better coding.
Qwen Cloud has just rolled out its Token Plan Individual, bringing unified credits starting at $6/month. The subscription spans Qwen, GLM, DeepSeek, and HappyHorse, plus early access to Qwen3.8-Max-Preview. Users are already raising questions about credit counting, API access, and payment errors.
Alibaba has just rolled out Qwen3.7-Plus, bringing vision and language into an API-first model for coding, GUI/CLI automation, and agent workflows. Early benchmarks and demos look promising, but third-party validation will be key.
OpenRouter has just rolled out Alibaba’s Qwen3.7-Max, positioning it as the flagship Qwen3.7 model for agent-centric coding, productivity, and long-horizon execution. The launch highlights claimed benchmark gains over Qwen3.6 and explicit prompt caching, as users press for more proof.
Hugging Face cofounder Julien Chaumond says running Qwen3.6 27B locally via Llama.cpp in Pi felt “pretty magical,” nearing Claude Opus for real code tasks. Replies quickly honed in on RAM, speed, and battery life trade-offs.
Alibaba’s Qwen team has unveiled Qwen3.6-Max-Preview as an early look at its next flagship model. The pitch: stronger agentic coding, improved instruction following, and better “real-world” reliability—alongside hints that more Qwen3.6 models are coming.