Skip to content
NewsModel release

Thinking Machines Lab releases Inkling, a 975B open-weight MoE with 1M-token context and native audio

· by Pondero Newsdesk

The short version

Mira Murati's Thinking Machines Lab released Inkling on July 15, 2026: a 975-billion-parameter mixture-of-experts model with 41B active parameters per forward pass, a 1-million-token context window, and native audio understanding. Full weights are publicly downloadable.

Thinking Machines Lab releases Inkling, a 975B open-weight MoE with 1M-token context and native audio

A 975-billion-parameter mixture-of-experts model that activates only 41 billion parameters per forward pass runs at roughly the inference cost of a dense 40B model. Native audio pretraining eliminates the transcription-pipeline dependency that limits audio workloads on most open-weight alternatives. Thinking Machines Lab released the full weights of Inkling on July 15, 2026; the model holds a 1-million-token context window and reasons natively over text, images, and audio.

What

Inkling's architecture routes each token through a subset of its expert layers rather than the full parameter set, per Thinking Machines Lab's announcement. The result is that a single forward pass draws on 41 billion active parameters despite a 975-billion total count. That gap matters for inference cost: the effective hardware footprint is closer to a 40B dense model than a 975B one.

Pretraining covered 45 trillion tokens of text, images, audio, and video, per the lab. Inkling reasons natively over text, images, and audio without routing through a transcription intermediate. Controllable thinking effort, an adjustable "effort" parameter ranging toward 1.0 in the benchmark tables, lets callers trade token spend for deeper reasoning in agentic pipelines with mixed-complexity tasks, per the lab.

Full weights are on Hugging Face at thinkingmachines/Inkling, both as the original checkpoint and as an NVFP4 checkpoint for NVIDIA Blackwell systems. Inference is available through five API partners: TogetherAI, Fireworks, Modal, Databricks, and Baseten, per the announcement. Inference framework support ships via SGLang, vLLM, llama.cpp, and Hugging Face Transformers through integrations the lab built with third-party teams.

Alongside Inkling, Thinking Machines Lab released a preview of Inkling-Small: a 276-billion-parameter MoE with 12 billion active parameters. The full weights for Small are not yet available; the lab said it is finishing testing.

Why it matters

The lab's stated goal is customization over benchmark leadership, which shapes the fine-tuning decision for anyone evaluating open-weight frontier models. "Inkling is not the strongest overall model available today, open or closed," the announcement states.

For operators considering open-weight models for fine-tuning pipelines, that framing shapes the decision. Inkling competes not as a frontier inference model but as a base worth specializing. The 41B active parameter footprint makes it runnable on mid-tier cloud instances that could not serve a dense model at anything approaching 975B scale. Teams on Databricks or Baseten, or running self-hosted vLLM, can experiment with the full model at a cost that reflects the active parameter count, not the total.

The audio-native pretraining is the least obvious differentiator. Open-weight models at large scale typically handle audio through a transcription layer. Training Inkling directly on audio means fine-tuning jobs targeting voice, speech, or audio classification can work with the raw audio embedding space rather than a text approximation. That opens workloads that are awkward on transcription-dependent architectures, particularly real-time conversational agents and audio grading pipelines.

Benchmark picture

Thinking Machines Lab published self-reported benchmark scores comparing Inkling against Nemotron 3 Ultra, GLM 5.2, GPT 5.6 Sol, and Claude Fable 5, per the announcement. With effort set to 0.99:

  • HLE (text only): 29.7%; HLE with tools: 46.0%
  • AIME 2026: 97.1%
  • GPQA Diamond: 87.2%
  • SWEBench Verified: 77.6%; SWEBench ProPublic: 54.3%
  • Terminal Bench 2.1: 63.8%
  • MCP-Atlas (agentic general): 74.1%
  • MMMU Pro Standard 10 (vision): 73.5%
  • VoiceBench (audio): 91.4%
  • StrongREJECT (safety): 98.6%

Inkling-Small at the same effort level posted 29.6% on HLE text, 46.6% with tools, 95.1% on AIME 2026, and 88.3% on GPQA Diamond. The gap between 12B and 41B active parameters turns out small on reasoning tasks, per the lab's tables. On Tau 3 Banking (a real-world agentic banking task), Inkling scored 23.7% versus Inkling-Small's 13.6%, a larger separation that suggests the full model's advantage sharpens on complex, multi-step tasks.

These figures are self-reported by Thinking Machines Lab. No independent reproduction on LM Arena or METR had been published as of July 15, 2026.

Context

Thinking Machines Lab was founded in late 2024. Murati left OpenAI in September 2024 after serving as CTO; she had been a central figure in GPT-4 and DALL-E's development. The lab's team includes researchers from OpenAI, DeepMind, and Google Brain, per public reporting. Inkling is the lab's first model release.

The lab runs a separate fine-tuning and deployment platform called Tinker. Inkling is available for fine-tuning on Tinker today; the lab added a playground interface for chatting with the model before committing to a run. Pricing carries a 50% introductory discount, with full rates documented on the Tinker platform.

To demonstrate what customization looks like in practice, the lab published a self-fine-tuning demo: Inkling, running inside the OpenCode agentic harness, wrote its own fine-tuning job, trained a lipogram variant that avoids the letter "e," and loaded the updated weights into its own runtime in approximately 27 minutes. The exercise is a staged demonstration, not a standard deployment pattern, but it shows the feedback loop between the model and the Tinker platform operating end-to-end.

A Python package called tml-renderer ships alongside the weights for handling tool calls, reasoning content, and multimodal inputs reliably during sampling and post-training.

What to watch next

Independent benchmark reproduction on LM Arena and METR is the clearest near-term signal. A SWEBench Verified score of 77.6% and an AIME 2026 score of 97.1% would, if confirmed externally, put Inkling at the upper end of the open-weight coding and reasoning tier. The MCP-Atlas score of 74.1% is the figure most directly relevant to teams building agentic tool-use workflows.

Inkling-Small's full weight release is the near-term product signal to watch: 12B active parameters with benchmark performance close to the larger model is the configuration that makes deployment practical for the widest range of teams. The lab gave no release date.

Sources