Thinking Machines has released Inkling, a general-purpose multimodal model with 975B total parameters (41B active), a 1M token context window, and native text, image, and audio input support. Modal partnered with Thinking Machines for day-zero availability as a Managed Endpoint with token-based pricing. Inkling uses a mixture-of-experts architecture with a local attention layout — five sliding window attention layers for every one global attention layer — enabling high compute efficiency. Modal further accelerated inference using a custom DFlash speculator tuned to Inkling's architecture, achieving 250 tokens/second per user at 2.5M TPM on 8x B200 GPUs, 67% faster than the model's built-in speculative path. The DFlash speculator uses block diffusion drafting with all-local attention and causal layers for better kernel support, keeping drafter cost flat as blocks grow.
