Mira Murati's Thinking Machines Drops Inkling, a 975B Open Multimodal Reasoning Model

Thinking Machines Lab just released Inkling, its first model trained from scratch with full weights publicly available. Inkling reasons natively over text, images, and audio, balances cost against performance with variable and efficient thinking effort, and is trained to exhibit safe behavior across modalities. This is the first open-weights model from the lab founded by former OpenAI CTO Mira Murati and co-founder John Schulman, and it arrives with a full ecosystem of inference and fine-tuning support baked in from day one.

A big model, built broad

Inkling is a mixture-of-experts transformer with 975B total parameters and 41B active parameters. It supports a context window of up to 1M tokens and was pretrained on 45 trillion tokens of text, images, audio, and video. The MoE design means only a fraction of the model activates per token, keeping inference costs manageable despite the massive total parameter count.

Inkling is not the most performant model available today, closed or open. The team trained it for solid capabilities across the board rather than state-of-the-art performance in a single area, to serve as a foundation for the models they will train in the future. That framing matters: this is a platform play, not a benchmark chase.

The audio story is the real headline

Most multimodal models treat audio as an afterthought, bolting on a speech-to-text layer before the LLM ever sees it. Inkling takes a different approach. The multimodal components are trained from scratch on general-domain data, using an encoder-free architecture for audio and vision inputs. Audio signals are input as discrete dMel spectrograms, while images are encoded as patches of 40x40 pixels using a four-layer hMLP. Both are transformed via a lightweight embedding layer and processed jointly with text tokens.

The results on audio benchmarks are notable for an open-weights model:

• VoiceBench: 91.4% (vs. 94.3% for Gemini 3.1 Pro and 88.8% for Qwen3-Omni)
• MMAU: 77.2% (vs. 82.5% for Gemini 3.1 Pro and 77.5% for Qwen3-Omni)
• AudioMC: 56.6% (vs. 66.8% for Gemini 3.1 Pro, and 24.3% for Qwen3-Omni)

On AudioMC in particular, Inkling more than doubles Qwen3-Omni's score, placing it in a different tier among open-weights models for audio understanding.

Controllable thinking: pick your spot on the cost curve

One of Inkling's most practically useful features is what the team calls controllable thinking effort. Rather than always running at full reasoning depth, you can dial the effort level between 0.2 and 0.99, trading token count for performance. Inkling spends roughly 1/3 as many tokens to achieve the same performance as Nemotron 3 Ultra on Terminal Bench 2.1. For workloads where you're running millions of inferences, that's a significant cost lever.

This isn't just a sampling temperature trick. The team specified the model's effort level on different samples by changing the system message and adjusting the per-token cost during RL training. This caused the model to use a different amount of tokens in different rollouts and learn the ability to control thinking effort.

How it was built

The MoE design largely follows DeepSeek-V3. Each MoE layer contains 256 routed experts and 2 shared experts, with 6 routed experts active per token. A few architectural choices stand out from the standard recipe:

• Attention layers interleave sliding-window and global layers at a 5:1 ratio with 8 KV heads.
• Position is encoded with a relative positional embedding rather than the more widely adopted Rotary Positional Embedding (RoPE), which the team found performs better and extrapolates better to longer sequences.
• Short convolutions are applied at two points: after the key and value projections in each attention layer, and on the attention and MLP residual branch outputs before they rejoin the main residual stream.

Inkling was pretrained with a hybrid optimization strategy: Muon for the large matrix weights and Adam for other parameters, using a coupling between weight decay and learning rate that kept the overall size of the model's weights stable throughout training.

Post-training was bootstrapped with an initial SFT on synthetic data generated by open-weight models including Kimi K2.5. The bootstrap accounts for a small fraction of compute, with the majority being employed for large-scale RL on synthetic and human-created environments.

RL at scale, and a surprising side effect

The team scaled RL to over 30M rollouts, with stable training sustained over just two continuous runs. Reasoning performance improved log-linearly throughout the entire process, resulting in a significant increase overall. The aggregate reasoning eval score moved from 0.264 at SFT initialization to 0.356 at the released checkpoint.

Something unexpected happened during this process. An emergent shift in reasoning style occurred over the course of RL training. The chain of thought became more concise over time, dropping grammatical overhead while remaining comprehensible and leaving the final response unaffected. Early in training, the model writes full grammatical sentences in its scratchpad. Late in training, it drops articles and connectives, producing a compressed but still-readable trace. This chain-of-thought condensation was independently observed by the Cognition team during SWE-1.7 training, suggesting it may be a general property of large-scale RL rather than something specific to Inkling.

Calibration as a first-class feature

Beyond raw capability, Thinking Machines invested heavily in what they call epistemics: the model's ability to express the right level of confidence, follow instructions precisely, and resist censorship pressure. Training calibration used RL against proper scoring rules on a large corpus of resolved real-world questions. On ForecastBench (no search), Inkling scores 61.1, competitive with Gemini 3.1 Pro (61.1) and above Claude Opus 4.8 (54.6) and GPT-5.5 (59.1).

Two automated graders were used for instruction following: a rubric grader that scores responses against a checklist of what a good answer should contain, and a claims grader that verifies each factual claim in the response using agentic web search. Together, the two graders improve helpfulness and reduce hallucination at the same time, rather than trading one for the other.

Where it falls short

Inkling is competitive but not at the top of the leaderboard across the board. A few areas where it trails:

• Hard reasoning (HLE text-only): 30.0% vs. 40.1% for GLM 5.2 and 53.3% for Claude Fable 5
• Factuality (SimpleQA): 43.9% vs. 57.0% for DeepSeek V4 Pro and 77.3% for Gemini 3.1 Pro
• Terminal Bench 2.1: 63.8% vs. 82.7% for GLM 5.2 and 91.9% for GPT 5.6 Sol

Inkling may exhibit general limitations common to foundation models, including hallucination, occasional failures to follow instructions precisely, and degraded performance in long multi-turn conversations. The team is transparent that this is a foundation for future work, not a claim to be the best model available.

Inkling-Small is already in preview

Alongside Inkling, the team is sharing a preview of Inkling-Small, a 276B parameter (12B active) mixture-of-experts model with a different performance/latency trade-off. Inkling-Small matches or exceeds its larger sibling on many benchmarks, the result of improvements made to the pre-training data mix and recipe for the smaller model.

With 12B active parameters and controllable thinking effort, Inkling-Small is a natural fit for workloads where cost and latency matter, such as coding, using LLMs to grade, or generating synthetic data for other models. Full weights for Inkling-Small are coming once testing is complete.

How to get it

Inkling is available on Tinker today with context length options of 64K and 256K tokens. It is offered at a 50% discount for a limited time. The Inkling Playground in the Tinker console lets you chat with the model for free while you evaluate it.

Full weights are on Hugging Face under an Apache 2.0 license, in both BF16 and NVFP4 (optimized for NVIDIA Blackwell). For local deployment, the model is supported across:

• SGLang and vLLM (via community PRs from RadixArk and Inferact)
• llama.cpp via Unsloth
• TokenSpeed via Lightseek
• Hugging Face Transformers

API access is available through Together, Fireworks, Modal, Databricks, and Baseten. The Tinker Cookbook has been updated with three new recipes specifically showcasing Inkling's audio capabilities.

The broader significance here is less about any single benchmark and more about what Thinking Machines is building toward. Inkling is designed to serve as the background reasoning model in the interaction models system the company recently introduced, which enables users to collaborate naturally using voice and vision in real-time. The open weights, the fine-tuning platform, and the audio-first design are all pieces of the same bet: that the next generation of useful AI will be customizable, multimodal, and owned by the people who deploy it.