---
title: "Together AI brings Thinking Machines Lab’s new model Inkling on day 0"
description: "Thinking Machines Lab released Inkling, and Together AI announced day-zero access through its production inference platform. The multimodal mixture-of-experts model accepts text, image, and audio inputs and generates text through a unified decoder architecture, with controllable inference effort for trading reasoning depth, token usage, and latency. Inkling has 975B total parameters, 40B active parameters per token, a 1M-token context window, query-conditioned relative attention, short causal convolutions, and shared-sink MoE routing; Together serves it with a FlashAttention-4-based kernel. Preliminary evaluations report strong results across scientific reasoning, mathematics, coding, agentic, vision, and audio benchmarks, while the model is also post-trained for forecasting and calibrated prediction. It is available on Together Serverless through OpenAI-compatible APIs without infrastructure provisioning."
---

# Together AI brings Thinking Machines Lab’s new model Inkling on day 0

[Together AI](https://yomu.fyi/company/together-ai) · Jul 15, 2026

**Type:** Announcement

## Summary

Thinking Machines Lab released Inkling, and Together AI announced day-zero access through its production inference platform. The multimodal mixture-of-experts model accepts text, image, and audio inputs and generates text through a unified decoder architecture, with controllable inference effort for trading reasoning depth, token usage, and latency. Inkling has 975B total parameters, 40B active parameters per token, a 1M-token context window, query-conditioned relative attention, short causal convolutions, and shared-sink MoE routing; Together serves it with a FlashAttention-4-based kernel. Preliminary evaluations report strong results across scientific reasoning, mathematics, coding, agentic, vision, and audio benchmarks, while the model is also post-trained for forecasting and calibrated prediction. It is available on Together Serverless through OpenAI-compatible APIs without infrastructure provisioning.

## Context

Inkling is presented as a new model from Thinking Machines Lab, with Together AI providing day-zero access through its Serverless production inference platform. The announcement focuses on supporting Inkling’s multimodal inputs, controllable reasoning, large context window, and specialized attention mechanism in a managed serving environment.

## Approach / What changed

Together AI exposes Inkling through Serverless and OpenAI-compatible APIs, using an optimized FlashAttention-4-based attention kernel for its query-conditioned relative attention mechanism. The platform accepts text, image, and audio inputs through a unified API call, supports request-level inference-effort control, and allows developers to scale from experimentation to dedicated production capacity.

## Takeaways

- Inkling accepts text, image, and audio inputs and processes them through one decoder stack, producing text outputs without separate modality-specific pipelines.
- The model combines query-conditioned relative attention, five local-attention layers followed by one full-attention layer, four-token causal convolutions, and shared-sink mixture-of-experts routing.
- Developers can adjust inference effort per request to balance reasoning depth, token usage, and latency, while Together Serverless handles infrastructure and GPU provisioning.

## FAQ

### What kinds of inputs and outputs does Inkling support?

Inkling accepts text, image, and audio inputs. It processes all three through the same decoder stack and generates text outputs.

### How does Inkling make reasoning effort configurable?

Inkling provides an adjustable inference-effort setting. Developers can use it to trade off reasoning depth, latency, and token spend for different workloads.

### What makes Inkling's architecture different from a conventional decoder-only Transformer?

Inkling uses query-conditioned relative attention, sliding-window and full causal attention, short causal convolutions, and a mixture-of-experts feed-forward design with shared expert sinks. Its multimodal embedding towers insert image and audio representations into the same input sequence as text tokens.

**Tags:** [AI](https://yomu.fyi/topic/ai), [Inference](https://yomu.fyi/topic/inference), [Mixture of Experts](https://yomu.fyi/topic/mixture-of-experts), [Multimodal](https://yomu.fyi/topic/multimodal)

- Source: [Together AI](https://www.together.ai/blog/together-ai-brings-thinking-machines-labs-new-model-inkling-on-day-0)
- Source URL: https://www.together.ai/blog/together-ai-brings-thinking-machines-labs-new-model-inkling-on-day-0
- Ingested by Yomu: 2026-09-19T10:13:40.595Z

[Read original post](https://www.together.ai/blog/together-ai-brings-thinking-machines-labs-new-model-inkling-on-day-0)