---
title: "Inkling by Thinking Machines now available on Modal"
description: "Thinking Machines released Inkling, a general-purpose multimodal model that accepts text, image and audio inputs and generates text outputs. The 975B-parameter mixture-of-experts transformer activates 41B parameters per token, uses a 1M-token context window, and employs local attention: five sliding window layers for every global layer. Modal provided day-zero support with a custom DFlash block-diffusion speculator, achieving 250 tokens per second per user on 8x B200s with 67% faster interactivity than the model’s built-in multi-token prediction path. The DFlash drafter uses all local attention and causal layers, keeping cost flat as block size grows. Inkling is available today as an OpenAI-compatible shared endpoint through Modal Auto Endpoints with token-based pricing, allowing users to try the model interactively."
---

# Inkling by Thinking Machines now available on Modal

[Modal](https://yomu.fyi/company/modal) · Jul 15, 2026

**Type:** Announcement

## Summary

Thinking Machines released Inkling, a general-purpose multimodal model that accepts text, image and audio inputs and generates text outputs. The 975B-parameter mixture-of-experts transformer activates 41B parameters per token, uses a 1M-token context window, and employs local attention: five sliding window layers for every global layer. Modal provided day-zero support with a custom DFlash block-diffusion speculator, achieving 250 tokens per second per user on 8x B200s with 67% faster interactivity than the model’s built-in multi-token prediction path. The DFlash drafter uses all local attention and causal layers, keeping cost flat as block size grows. Inkling is available today as an OpenAI-compatible shared endpoint through Modal Auto Endpoints with token-based pricing, allowing users to try the model interactively.

## Approach / What changed

Modal developed a custom DFlash block-diffusion speculator adapted to Inkling’s local attention architecture, using all-local attention and causal layers to achieve faster drafting with flat cost scaling.

## Takeaways

- Inkling is a 975B-parameter mixture-of-experts transformer with 41B active parameters, a 1M-token context window, and a local attention layout with five sliding window layers per global layer.
- On 8x B200s with Modal’s DFlash speculator, Inkling reached 250 tokens per second per user and 67% faster interactivity than its built-in multi-token prediction path at matched throughput.
- The DFlash drafter uses all local attention and causal layers, enabling flat cost scaling as block size increases, which suits models shifting toward local attention.

## FAQ

### How does Inkling's local attention work?

For every six attention layers, five use sliding window attention over recent tokens and one uses full attention over the whole sequence. This biases compute toward recent tokens for greater efficiency.

### What performance does DFlash achieve on Inkling?

On 8x B200s, Modal's DFlash speculator achieved 250 tokens per second per user at 2.5M TPM per-GPU throughput, with 67% faster interactivity than the model's built-in speculative path.

### How does DFlash differ from MTP?

Multi-token prediction heads draft one token at a time, causing performance ceilings. DFlash's block-diffusion drafter generates a block in one parallel pass, conditioned on target features, with cost staying flat as block size grows.

**Tags:** [DFlash](https://yomu.fyi/topic/dflash), [Inkling](https://yomu.fyi/topic/inkling), [Mixture of Experts](https://yomu.fyi/topic/mixture-of-experts), [Modal](https://yomu.fyi/topic/modal), [Speculative Decoding](https://yomu.fyi/topic/speculative-decoding)

- Source: [Modal](https://modal.com/blog/inkling-by-thinking-machines-labs-now-available-on-modal)
- Source URL: https://modal.com/blog/inkling-by-thinking-machines-labs-now-available-on-modal
- Ingested by Yomu: 2026-09-20T07:47:12.396Z

[Read original post](https://modal.com/blog/inkling-by-thinking-machines-labs-now-available-on-modal)