DEVOURED
Introducing Inkling 2 Small
Thinking Machines released Inkling 2 Small, a 594B-parameter Mixture-of-Experts model that excels at agentic coding and visual reasoning.
What: Inkling 2 Small is a Mixture-of-Experts model with 17B active parameters, offering visual-feedback-based iteration for tasks like website generation and 3D design. It achieved high performance on agentic coding benchmarks and includes vision-in-the-loop capabilities for complex task execution.
Why it matters: The focus on 'agentic' models that can write, execute, and fix their own code through environmental feedback marks a shift from passive text generation to autonomous workspace execution.
Takeaway: Test Inkling 2 Small via the Tinker Playground to evaluate its agentic capabilities for your own development workflows.
Deep dive
- Features a 292M-parameter vision encoder and supports image inputs.
- Uses RL-based training on hundreds of thousands of long-horizon coding tasks.
- Implements an asynchronous monitor to inspect tool calls and prevent model misbehavior during training.
- Achieves high performance on benchmarks like DesignArena and SWEBench Pro.
- Utilizes NVFP4, a 4-bit floating-point format native to NVIDIA Blackwell GPUs, during training.
- Employs sticky routing and DRAM KV-cache offloading to maintain high prefix cache hit rates for long-context tasks.
Decoder
- Mixture-of-Experts (MoE): A model architecture where only a subset of parameters ('experts') is activated for each input, improving inference speed and efficiency.
- Agentic: A model characteristic where the system can act autonomously, use tools, and iterate on solutions through a feedback loop.
- KV-Cache: A memory buffer that stores key-value pairs of previously generated tokens to speed up inference by avoiding redundant computation.
Original article
Full article content is not available for inline reading.
Explore all 54 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi