
开发者要基座模型,蚂蚁真给了!6大训练关键节点同步开源 (opens in a new tab)
百灵为开发者提供了多个可继续训练的起点。今日,蚂蚁百灵开源Ling-3.0-tiny-base和Ling-3.0-flash-base。除最终Base模型权重外,团队还同步开放了两款模型的预训练和中期训练权重,共计6个checkpoints。
Total 235 articles

百灵为开发者提供了多个可继续训练的起点。今日,蚂蚁百灵开源Ling-3.0-tiny-base和Ling-3.0-flash-base。除最终Base模型权重外,团队还同步开放了两款模型的预训练和中期训练权重,共计6个checkpoints。

RoboRisk evaluates model-controlled robots on 92 simulated and thirteen physical tasks, testing LEASH’s three gates. Some configurations incurred no harm in the physical sample; this does not guarantee safety in open environments.

Sage-S2.1 Flash supports scientific data and demonstrates 3D Slicer, Blender, QGIS and CAD workflows. Its team reports 90.3 on MLE-bench Lite; rankings apply to the listed models and evaluation settings.

FlashStates manages recurrent state and KV caches together for hybrids using KDA, GDN and Mamba-style layers. The team tests prefix reuse, prefill speed and tiered-cache benefits on B200 GPUs with long-context workloads.

Philipp Schmid describes an agent using Bash and a media viewer. Across over 350 tasks, harness scores were close while tokens and costs differed. Each task ran once, so score differences still require repeated experiments.

QbitAI summarizes Latent Space's 30-question speed test and solar-system, growing-tree and ray-tracing demos, reporting peak throughput in that setup and differences from two Flash models.

The team post-trains GLM-5.3-Flash using task filtering, asynchronous sampling, and a shared runtime framework, reporting terminal and software engineering benchmark results and estimated costs.

团队以 Qwen3.8-Flash 为基础进行后训练,主模型 1250 亿参数、激活 60 亿。厂商评测中部分 Agent 任务分数提高,计算与数学类别下降,文章也给出四个官方案例。

FlashProve 通过权重承诺和零知识证明验证受支持的推理结果。在 Llama 4 Maverick 实验中,每份证明检查一个 Token,CPU 多线程验算约 17 秒,服务方需要参与生成证明。

FlashProve 用权重承诺与零知识证明探索模型推理验证。团队在400B参数的Llama 4 Maverick上报告约110 Token/s证明吞吐;当前实验为1024 Token窗口中的单个输出Token,仍受整数计算、贪婪解码及较高证明成本限制。