Realtime AI News
Qwen Releases Open Multimodal Model Qwen3.8-27B with FP8 Variant on Hugging Face
Qwen has published Qwen3.8-27B, a new image-text-to-text multimodal model, on its official Hugging Face registry under the Apache 2.0 license. The same day, the team also released an FP8 quantized variant, Qwen3.8-27B-FP8, to lower the barrier to deployment.
Qwen has updated its official model registry on Hugging Face, publishing Qwen3.8-27B, a new multimodal model built around an image-text-to-text pipeline with conversational support.
According to the model card, Qwen3.8-27B is built on the transformers library, ships weights in safetensors format, carries the permissive Apache 2.0 license, and is marked endpoints_compatible for easier integration into inference services.
The same day, the team also released an FP8 quantized variant, Qwen3.8-27B-FP8, based on the original model and also licensed under Apache 2.0. Quantized versions typically reduce memory footprint and inference cost, making them attractive for production deployments.
Registry data shows the main model page already has 7,818 likes, signaling strong community interest, while the freshly uploaded FP8 version is still at the very start of its download curve.
At 27B parameters, the model sits in a practical middle band for open multimodal models, balancing capability against deployment requirements, and the Apache 2.0 license lets developers integrate or fine-tune it freely in commercial products.
The bigger signal is that the open multimodal race keeps accelerating, and leading teams now ship quantization variants alongside the main release, treating inference efficiency and ease of deployment as standard parts of a model launch.
Next, watch for the detailed technical documentation and benchmark results that usually follow a model card, community evaluations, and how quickly inference frameworks in the Hugging Face ecosystem add support for Qwen3.8-27B.
Why it matters
By open-sourcing Qwen3.8-27B under Apache 2.0 and pairing it with an FP8 variant on day one, Qwen lowers the barrier to building multimodal applications and raises the competitive bar for open-source multimodal models.
Nearby Updates
All08/14, 22:50
French Startup Kog Bets GPUs Are Not the Wrong Fit for Agentic Inference
French startup Kog argues that the idea of GPUs being poorly suited to agentic workloads may be a misconception, and says it is working to squeeze more inference out of GPUs. TechCrunch's profile shows the company is betting on deep, hardware-level optimization rather than dedicated chips.
08/14, 23:43
Meta releases open-weight model Glimmer as Zuckerberg argues AI should be 'for everyone'
Meta released Glimmer this week, an open-weight AI model anyone can download and run on their own hardware, while its more powerful Muse Spark model stays locked behind Meta's own APIs. The release arrived alongside a letter from Mark Zuckerberg arguing AI should be 'for everyone,' a claim TechCrunch's Equity show questions.
08/14, 22:05
New Forecast Warns Natural Gas Prices Could Triple, Hitting Hyperscaler AI Data Center Bills
A new forecast cited by TechCrunch warns that natural gas prices could triple in parts of the United States, potentially saddling hyperscalers with massive power bills for their AI data centers. If the forecast holds, energy costs will become one of the hardest variables in the AI compute buildout.
08/14, 18:25
Teco Opens Recruitment for AI Accelerator Card Model Adaptation Track at National AI+Education Competition
The national AI+Education Innovation Application Skills Competition, organized by the China Association for Educational Technology, is underway, and strategic partner Teco has opened recruitment for its AI + Accelerator Card Model Adaptation Track under the university student category. Teams will adapt and optimize models on Teco's self-developed AI accelerator cards with free cloud resources, with registration and code submission closing on October 15, 2026.