Mira Murati's Inkling Tops Every Open-Weight Model on ARC-AGI-2

Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, just got its first major benchmark crown. ARC Prize has officially verified that Inkling, Thinking Machines' debut open-weight model, is now the highest-scoring open-weight model on both ARC-AGI-1 and ARC-AGI-2. It hits 79.5% on ARC-AGI-1 and 36.5% on ARC-AGI-2, at costs of $0.30 and $0.64 per task respectively. For context, that ARC-AGI-2 score puts it ahead of every other open-weight model that has been formally evaluated by ARC Prize, in a benchmark where frontier closed models like GPT-5.5 sit around 85%.

What ARC-AGI actually tests

ARC-AGI is not a typical benchmark. Each puzzle ships with 2-8 examples, and the model must infer a hidden rule (color swaps, symmetry, recursion, compositional transforms) and apply it to a withheld test grid, with no memorized facts, just pattern discovery and reasoning. It is deliberately designed to resist the kind of pattern-matching that large training corpora enable. Average individual human performance on ARC-AGI-2 is 66%. The benchmark is graded with zero tolerance: answers must match ground-truth output exactly, any discrepancy in length, structure, or formatting fails the task, and no partial credit is given, with attempts limited to three per test input.

ARC-AGI-2 launched in early 2025 with every frontier model at 0%. Progress over the past year has been steep. The open-weight side of the leaderboard, however, has lagged far behind closed-weight systems, making Inkling's 36.5% a meaningful data point for the open-source community.

What Inkling actually is

Thinking Machines Lab released its first in-house model, and unlike the flagships from OpenAI, Anthropic, or Google, anyone can download it. Inkling is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active, a 1 million token context window, and 45 trillion tokens of pretraining across text, images, audio and video. The MoE design (Mixture-of-Experts) means the model routes each token through a small subset of its parameters, keeping inference fast and cheap despite the massive total parameter count.

Thinking Machines trained it to be a broad, balanced foundation model: strong across many domains, flexible enough to adapt. Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning. That is an unusually candid framing for a model launch, and it sets up a different value proposition than the usual benchmark race.

The architecture and training story

The Mixture-of-Experts transformer boasts 975 billion total parameters, with 41 billion active, and supports a context window of up to 1 million tokens. Pretrained on 45 trillion tokens of text, images, audio, and video, Inkling represents the first in a new family of models. The MoE design largely follows DeepSeek-V3, with 256 routed experts and 2 shared experts per layer, and 6 routed experts active per token. Notably, the team used a relative positional embedding instead of the more common RoPE, claiming better extrapolation to long sequences.

The post-training story is where things get interesting. The team ran large-scale asynchronous RL (reinforcement learning) over more than 30 million rollouts, with reasoning performance improving log-linearly throughout. One emergent behavior stood out: the chain of thought became progressively more compressed and telegraphic over the course of RL, without being explicitly rewarded for it. The model dropped grammatical overhead while staying coherent, a compression effect also recently observed by the Cognition team in their SWE-1.7 training.

Multimodality was baked in from scratch rather than added as an afterthought. Thinking Machines opted for an encoder-free architecture for audio and vision inputs. Audio signals are input as 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.

What it is good at

Inkling's benchmark profile is deliberately broad. Key numbers at full effort (effort=0.99):

• Math: 97.1% on AIME 2026, 87.2% on GPQA Diamond
• Agentic coding: 77.6% on SWE-bench Verified, 54.3% on SWE-bench Pro Public, and 63.8% on Terminal Bench 2.1
• Instruction following: 79.8% on IFBench, ahead of every other model in the comparison table including Claude Fable 5 and GPT 5.6 Sol
• Audio: 56.6% on Audio MC, 77.2% on MMAU, and 91.4% on VoiceBench, placing it among the strongest open-weights audio models
• ARC-AGI (verified): 79.5% on ARC-AGI-1, 36.5% on ARC-AGI-2 , top open-weight on both

A standout feature is controllable thinking effort. This allows developers to balance performance with token efficiency, which is critical for applications where cost and latency are binding constraints, enabling collaborative and iterative development. In practice, Inkling matches Nemotron 3 Ultra on Terminal Bench 2.1 at roughly a third of the tokens.

Where it falls short

The honest benchmarks tell the story. On factuality, Inkling scores 43.9% on SimpleQA Verified, well behind closed-weight models like Gemini 3.1 Pro (77.3%) and Claude Fable 5 (68.3%). On advanced reasoning, its 29.7% text-only HLE score trails GLM 5.2 (40.1%) and Claude Fable 5 (53.3%). ARC-AGI-3, the newest and hardest version of the benchmark, has not yet been evaluated. ARC Prize noted that ARC-AGI-3 evaluations are more operationally intensive, so results will roll out over the next few weeks.

There is also a gap relative to the closed-weight frontier on ARC-AGI-2 itself. As of mid-July 2026, GPT-5.5 leads the ARC-AGI-2 leaderboard with 85%, followed by GPT-5.4 Pro at 83.3% and Gemini 3.1 Pro at 77.1%. Inkling's 36.5% is a strong open-weight result, but the gap to the frontier remains large.

How to use it and what it costs

Inkling's weights can be downloaded through Hugging Face under an Apache 2.0 license, while the model can also be fine-tuned through Thinking Machines' Tinker training platform. Thinking Machines is offering Inkling at a 50% discount for a limited time on Tinker, with context length options of 64K and 256K tokens. The Inkling Playground in the Tinker console is free for a limited time.

For inference, the ecosystem support is broad. The full weights are on Hugging Face, both as the original checkpoint and as an NVFP4 checkpoint for efficient inference on NVIDIA Blackwell systems. API access is available via TogetherAI, Fireworks, Modal, Databricks, and Baseten. Inference framework support covers SGLang, vLLM, llama.cpp, and Hugging Face Transformers.

For reproducibility, ARC Prize's open benchmarking repo lets anyone reproduce the public evaluation results.

The smaller sibling worth watching

A lighter preview, Inkling-Small, runs 276B parameters with 12B active and matches or beats its larger sibling on several benchmarks, including IFBench at 83.4% against 79.8%. With 12B active parameters and controllable thinking effort, it is a natural fit for workloads where cost and latency matter, such as coding, LLM-as-grader pipelines, or synthetic data generation. Full weights for Inkling-Small are pending completion of testing.

Why this matters beyond the headline

Inkling offers a different proposition: that the future of advanced AI may be shaped not only by whoever trains the most powerful general-purpose model, but also by who gives developers the strongest foundation for building models of their own. The ARC-AGI result is the first independent, third-party verification of that claim. It shows that an open-weight model trained from scratch, with full weights available for fine-tuning, can reach competitive abstract reasoning performance without being a closed-API system.

The ARC-AGI-3 evaluation is still pending for Inkling. Given that ARC-AGI-3's launch humbled every frontier model back to near-zero, that result will be the real test of whether Inkling's reasoning capabilities generalize to the next level of difficulty. For now, the open-weight community has a new benchmark leader, and it comes with Apache 2.0 weights and a fine-tuning platform attached.