Unsloth
PulseAugur coverage of Unsloth — every cluster mentioning Unsloth across labs, papers, and developer communities, ranked by signal.
- 2026-05-19 product_launch Unsloth released version 0.1.41-beta with bug fixes and performance improvements. 来源
- 2026-05-19 product_launch Unsloth released version v0.1.405-beta with performance and feature enhancements. 来源
- 2026-05-06 product_launch Unsloth released a new API inference endpoint for local LLM deployment. 来源
- 2026-04-23 product_launch Unsloth released a beta update with a redesigned UI and new chat management features. 来源
- 2026-04-08 product_launch Unsloth released updates and fixes for the Gemma 4 model and its associated Studio product. 来源
7 天有情绪数据
-
Unsloth 修复 Gemma 4 训练和量化错误
Unsloth 为 Gemma 4 模型发布了重要的修复补丁，解决了最初并非由 Unsloth 引起但影响训练和量化的问题。这些更新解决了诸如梯度累积期间的损失爆炸和较大模型变体出现的索引错误等问题，确保 Gemma 4 训练现在能在 Unsloth 框架内正常运行。此次发布还包括了比其他设置更快的训练速度和更低的 VRAM 使用量优化，以及增强了 Unsloth Studio 对各种模型类型和任务能力的更新。
-
Google Gemma 4模型现已可在Unsloth上运行和训练
Google发布了Gemma 4，这是一套包含E2B、E4B、26B-A4B和31B的全新模型。这些模型现已兼容Unsloth，一个优化模型训练和推理的平台。Unsloth使用户能够在仅6GB内存的设备上运行较小的Gemma 4模型，使其在手机等设备上可用，而较大的模型则需要约18GB内存。此次更新还显著提高了工具调用（tool calling）的准确性和稳定性，减少了错误并增加了允许的调用次数。
-
Alibaba's Qwen3.5-397B-A17B model offers multimodal capabilities and efficient inference
Alibaba has released Qwen3.5-397B-A17B, an open-weight, natively multimodal model featuring a hybrid attention mechanism and sparse Mixture-of-Experts architecture. The model boasts support for 201 languages and demonst…