Thinking Machines Lab Unveils deepseek-Inspired Inkling XCXWBN
benchmarks deepseek gpu
| Source: Mastodon | Original article
Thinking Machines Lab launches Inkling, a 975-billion-parameter AI model. It features a DeepSeek-inspired design and significant GPU requirements.
Thinking Machines Lab has launched Inkling, a 975-billion-parameter open-weight AI model inspired by DeepSeek's design. This new model boasts a massive parameter count and requires 2 TB of GPU power, with mixed results in benchmarks. Inkling's design is based on the MoE architecture, similar to DeepSeek-V3, with a sigmoid-based router and load-balancing bias.
The release of Inkling matters because it challenges existing models like Kimi and Nemotron, outperforming Nvidia's Nemotron 3 Ultra in several evaluations. This puts Inkling among the most capable open models available outside China, giving developers a new alternative in a market dominated by DeepSeek, Qwen, and Kimi. As an open-weights model under the Apache 2.0 license, Inkling is trained from scratch, offering a fresh perspective in the AI landscape.
As the AI landscape continues to evolve, it will be interesting to watch how Inkling performs in various tasks and how it compares to other models. With its impressive parameter count and open-weights design, Inkling has the potential to drive innovation and advancements in the field. Developers and researchers will likely be keen to explore Inkling's capabilities and limitations, and its impact on the AI market will be worth monitoring in the coming months.
Sources
Back to AIPULSEN