Inkling: What the New Open-Weights Model Can Do

By Alec Chizhik July 16, 2026 5 min read

On 15 July 2026, Thinking Machines Lab released Inkling, an open-weights MoE model under Apache 2.0. It processes text, images and audio, generates text and, according to the vendor, offers a context window of up to one million tokens. The weights are hosted on Hugging Face, and fine-tuning is possible via Tinker. The vendor explicitly does not position Inkling as the single strongest model overall; for enterprises, adaptability, infrastructure, cost and in-house testing count for more than any headline ranking.

Key Takeaways

Open Weights under Apache 2.0. Inkling comprises, according to the vendor, 975 billion total parameters and 41 billion active parameters.
Text, images and audio as input. The output is text; the maximum context window is stated as one million tokens. Weights are hosted on Hugging Face; fine-tuning is possible via Tinker.
Your own tests remain mandatory. Vendor benchmarks are vendor claims. What matters is evaluation within your own stack, cost and the right deployment option.

What Inkling brings to the table technically

What is Inkling? Inkling is an open-weights model from Thinking Machines Lab. The company releases the trained model weights under Apache 2.0, enabling teams to inspect them, integrate them into their own systems and continue training for defined tasks.

Inkling follows the pattern of large Mixture-of-Experts architectures: the total number of parameters is 975 billion, while, according to the vendor, 41 billion are active per inference step. This separates the model’s storage footprint from its runtime load – relevant for procurement and architecture because licence, memory demand and compute demand are different levers.

As an open-weights model under Apache 2.0, weights and associated artefacts can be downloaded, inspected and integrated into your own pipelines under the licence terms. Hosting the weights on Hugging Face lowers the entry barrier for teams already using Hugging Face workflows, internal model catalogues or CI-driven evaluation pipelines.

Multimodality on the input side is broad: text, images and audio can flow jointly into the context. The output is text. For typical enterprise scenarios, this means logging, research across documents and screenshots, evaluation of support calls or combining specifications with UI screenshots. The vendor states a maximum context window of one million tokens. That opens the door to lengthy briefings, large ticket histories or consolidated project documentation in a single pass. Whether and how stably this limit holds in practice remains a matter for your own load tests, using real tokenisations and your chosen serving stacks.

Operation, Customization, and What the Manufacturer Itself Downplays

Thinking Machines Lab explicitly states that Inkling is not the strongest model overall. This is a rare but useful clarification: it decouples marketing from the question of whether the model actually delivers economic and qualitative value in your specific use case. Vendor benchmarks are vendor data. They can serve as a starting point for hypotheses, but they do not replace an independent evaluation on your own data, using your own prompts, and reflecting real latency, throughput, and safety requirements.

Fine-tuning is supported via Tinker. For companies looking to align domain language, internal formats, or compliance tone, this is the practical lever alongside pure prompting and retrieval. Open weights and a fine-tuning path mean teams can version models, run A/B tests, and define rollback paths – provided governance, data access, and audit trails are part of the plan.

From an infrastructure standpoint, options include self-hosting, managed GPU clusters, and hybrid setups. The 41 billion active parameters define the rough compute class per request, while the 975 billion total parameters determine storage and distribution needs. Anyone seriously planning to deploy Inkling should align quantization, expert parallelism, batching, and storage costs early with finance and platform teams. Apache 2.0 simplifies internal use and redistribution under the license’s standard terms, but legal clearance within your organization remains a separate step.

What Decision-Makers Should Check

For the business case, four factors outweigh any public ranking: adaptability to your data and processes, available infrastructure and operating model, ongoing inference and storage costs, and quality in your own evaluation. Open weights shift budget from pure API seats to GPU time, engineering, and monitoring. It’s not automatic, but it is planable if token prices, utilization, and queue times are modeled transparently.

Multimodal inputs raise integration effort: image and audio pipelines require clear data-protection and retention rules, especially for customer voices and internal screenshots. A large context window tempts teams to dump everything into a single prompt. A better approach is an architecture that combines retrieval, segmentation, and traceable sources so you can control both costs and hallucination risk.

Pragmatic entry point: a tight pilot with measurable tasks, a baseline against your current model, identical test sets for quality and cost, followed by a decision on fine-tuning via Tinker and on the serving path. That way, Inkling remains a tool with clear KPIs rather than a blanket platform bet.

Frequently Asked Questions

What is Inkling?

Inkling is an open-weights Mixture-of-Experts model released on 15 July 2026 by Thinking Machines Lab under Apache 2.0. It processes text, images, and audio and generates text.

How large is the model?

According to the vendor, Inkling comprises 975 billion parameters in total and 41 billion active parameters per inference step. The maximum context window is stated as one million tokens.

Where are the weights available, and how can I adapt the model?

The weights are hosted on Hugging Face. Fine-tuning is possible via Tinker. Both features ease adoption for teams with existing MLOps and evaluation workflows.

Does the vendor claim Inkling is the strongest model?

No. The vendor explicitly states that Inkling is not the strongest model overall. Public vendor benchmarks should be read as vendor data and supplemented with your own tests.

What should companies focus on?

Focus on adaptability, infrastructure options, ongoing costs, and your own evaluation in the target use case. These factors matter more for operations than any sweeping ranking claim.