EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·24d ago·5 min read
Shapelearn Qwen 3.8 27B (13.1 GB VRAM)
TL;DR
ShapeLearn Qwen 3.8 27B has been released, showcasing significant improvements over its Lite version, particularly in quality and speed across various GPUs. The full models outperform the Lite models, achieving higher aggregate scores and throughput metrics.
✦ Why It Matters
Evaluate your GPU capabilities and consider upgrading to the full ShapeLearn models for optimal performance.
Key Takeaways
- Utilize GPU-5 for maximum performance with ShapeLearn models.
- Consider GPU-4 for a balance of speed and memory efficiency.
- Implement speculative decoding for improved throughput.
- Benchmark your current models against ShapeLearn for potential upgrades.
Full Summary
ShapeLearn Qwen 3.8 27B, released on August 14, 2026, includes both full and Lite models, with the latter being a quicker, less resource-intensive version. The full ShapeLearn models have been benchmarked against the Lite models and competing quantized models, demonstrating superior performance.
Notably, the GPU-5 model achieved 99.63% of the BF16 aggregate benchmark score, while GPU-4 remains a competitive option at 98.72%. Speculative decoding techniques, MTP and DFlash2, enhance throughput across all models tested.
The results indicate that larger models yield higher aggregate scores, while smaller models provide better throughput, making them suitable for different use cases. The findings suggest that the full ShapeLearn models are positioned at the forefront of performance metrics in the current landscape.