Inkling Small FP8 Dynamic
Built for reasoning, coding, and multimodal agent workloads where you want Inkling Small’s capabilities with about half the GPU memory.
Explore additional models
Frequently
asked questions
Inkling Small FP8 Dynamic is Reflection AI’s 8-bit version of Thinking Machines Lab’s Inkling Small. It has the same mixture-of-experts architecture (276B total parameters, 12B active), the same 1M-token context, and the same text, image, and audio input. The only change is that its weights are stored in 8-bit precision, which roughly halves the GPU memory it needs. It ships under the Apache 2.0 license. It is available for high-performance inference via Self-Serve Deployments in Crusoe Intelligence Foundry.
The two models share the same architecture, parameter counts, context length, input types, and Apache 2.0 license. The difference is precision: this model stores the FP8 Dynamic weights in 8-bit precision, which roughly halves the GPU memory needed to serve the model, with a small accuracy tradeoff. Choose Inkling Small FP8 Dynamic when you want to serve on fewer GPUs, and Inkling Small when you need the more accurate outputs.
Yes. You can deploy Inkling Small FP8 Dynamic for high-performance inference via Self-Serve Deployments on any available GPU, with a profile tuned for throughput or responsiveness. Because it needs roughly half the GPU memory of Inkling Small, it fits on fewer GPUs, which lowers the cost of a dedicated deployment.
Yes. Storing the weights in 8-bit precision doesn’t change how you call the model, so Inkling Small FP8 Dynamic keeps the same thinking-effort settings as Inkling Small, from minimal to extra-high. Use low effort for fast, high-volume tasks and higher effort for hard reasoning and agentic coding.


%204.png)
