Introduction
On July 15, 2026, Thinking Machines Lab — the research company founded by former OpenAI CTO Mira Murati — released Inkling, its first model trained from scratch and its first open-weights release. Inkling is a 975B-parameter Mixture-of-Experts model (41B active per token) that reasons natively over text, images and audio, supports a context window of up to 1M tokens, and ships under the permissive Apache 2.0 license. Artificial Analysis called it the leading U.S. open-weights model at the time of release.
That framing is the news. Until now, teams wanting the strongest open weights have largely reached for Chinese models like DeepSeek V4 or Kimi K3. Inkling is the first competitive Western entry in that tier — and it comes from a lab whose only prior product was Tinker, a fine-tuning platform, not a model.
Full details are on the model page: Inkling.
What was announced
Inkling is positioned as a broad, balanced generalist rather than a benchmark specialist. The headline facts:
- Architecture: Mixture-of-Experts transformer, 975B total parameters, 41B active per token.
- Multimodality: Native text, image and audio input — reasoned over directly, not through bolted-on encoders. Output is text.
- Context: Up to 1M tokens for the open weights; 64K and 256K options on the Tinker API.
- Training: Pretrained on 45 trillion tokens of text, images, audio and video.
- Controllable thinking effort: A first-class dial to trade reasoning depth against latency and token cost.
- License: Apache 2.0, with full weights and an NVFP4 quantized checkpoint on Hugging Face.
Thinking Machines also shipped a preview of Inkling-Small — 276B total parameters, 12B active — trained with a similar recipe for lower cost and latency. Notably, it matches or beats the flagship on several evaluations.
The benchmarks
The numbers below are vendor-reported, from Thinking Machines' launch announcement, measured at a thinking effort of 0.99. As with any first-party table, the developer picked the evals and settings — treat them as an upper bound and validate on your own tasks.
| Benchmark | Inkling | Inkling-Small |
|---|---|---|
| AIME 2026 | 97.1% | 95.1% |
| GPQA Diamond | 87.2% | 88.3% |
| HLE (with tools) | 46.0% | 46.6% |
| SWE-bench Verified | 77.6% | 77.4% |
| Terminal-Bench 2.1 | 63.8% | 52.7% |
| IFBench | 79.8% | 83.4% |
| VoiceBench | 91.4% | 90.0% |
| FORTRESS (adversarial) | 78.0% | 75.6% |
Independently, Artificial Analysis scored Inkling 41 on its Intelligence Index (v4.1) — three points above Nvidia's Nemotron 3 Ultra (38), and well ahead of Gemma 4 31B (29) and gpt-oss-120b (24), which is what earns it the "leading U.S. open weights" label. (Index versions are not comparable, so the v4.1 tag matters.) On the GDPval-AA v2 agentic benchmark it posted an Elo of 1238, above Kimi K2.6 and DeepSeek v4 Flash, while spending only ~25K output tokens per task versus 37–43K for competing models — the efficiency case for the controllable-effort design.
Why it matters
Three things stand out beyond the scores.
A competitive Western open-weights model. For much of the past year, the open-weights frontier has been defined by Chinese labs. Inkling gives U.S. and European teams a strong, permissively licensed alternative they can self-host, inspect and fine-tune without data leaving their environment.
Efficiency as a design goal, not an afterthought. The controllable thinking-effort dial and the low token spend per task target the real cost of running agents, where every reasoning step is billed and latency compounds. Inkling-Small keeping pace with the flagship reinforces the point — the smaller model may be the better default for many workloads.
Day-one fine-tuning. Inkling launched wired into Thinking Machines' Tinker platform, with 64K and 256K context recipes and a limited-time 50% discount, so fine-tuning on proprietary data is available immediately rather than months later.
The caveats
It is not an unqualified win. Inkling is priced above several open competitors — roughly $1.87 input / $4.68 output per million tokens on hosted endpoints — which Artificial Analysis flags as above-average for the tier. Serving a 975B-parameter model, even at 41B active, needs real GPU memory; the NVFP4 checkpoint helps, but this is not a laptop model. The launch benchmarks are first-party and measured at high effort, independent replication was still emerging at release, and Thinking Machines did not publish a knowledge-cutoff date.
How to access it
- Self-host: Download the weights (original or NVFP4) from Hugging Face under Apache 2.0.
- Fine-tune: Use Tinker, with 64K and 256K context options.
- Hosted APIs: TogetherAI, Fireworks, Modal, Databricks and Baseten, each setting its own pricing and limits.
Conclusion
Inkling is a notable release less for topping any single leaderboard than for what it represents: a genuinely open, commercially usable, competitive model from a well-funded U.S. lab, built for cost-controlled agentic and multimodal work. Whether it displaces the incumbent open models will come down to independent evaluation and real serving economics — but the open-weights field now has a serious new Western entrant. For the full specifications, pricing tiers and benchmark table, see our Inkling model page, or browse the models catalog to compare it against the alternatives.