On July 15, 2026, Thinking Machines Lab released Inkling — its first open-weight foundation model. The headline numbers are striking: 975 billion total parameters, a 1-million-token context window, native multimodal reasoning across text, images, and audio, and an Apache 2.0 license. For SMBs building local AI infrastructure that keeps data under their own control, this is one of the most significant model releases of 2026.
What Inkling Is, Technically
Inkling is a Mixture-of-Experts (MoE) architecture. Of its 975 billion total parameters, only around 41 billion are activated per token generation step. The remaining parameters stay idle, like a large team of specialists where only the right experts are pulled into each task.
That architecture has a practical consequence: the actual compute per token is far lower than a dense model of equivalent total size. For comparison, Qwen 3 235B-A22B — widely reported as one of the strongest locally deployable open-weight models — activates around 22 billion parameters per step. Inkling activates nearly twice as many active parameters while still being an efficient-per-token MoE.
The model was pretrained on 45 trillion tokens of text, images, audio, and video. It processes all three modalities natively without a separate encoder layer, making it possible to pass an image, a voice recording, and a text prompt in a single call. The 1-million-token context window supports processing of long documents, entire codebases, or extended transcripts in a single request.
A notable design choice: Inkling includes a thinking-effort dial, adjustable from 0.2 (fast, lightweight) to 0.99 (deep, compute-intensive reasoning). Teams can tune the trade-off between response speed and quality depending on the task.
Apache 2.0: Why the License Is the Story
Open-weight models exist on a spectrum. Some carry non-commercial restrictions. Others allow commercial use but prohibit fine-tuning or redistribution. Apache 2.0 is among the most permissive licences available:
- Commercial use with no royalties or per-seat fees
- Fine-tuning on your own data without requiring approval from the model creator
- Self-hosted deployment with no obligation to report usage back to Thinking Machines Lab
- Distribution of modified versions, including internal products
Under our reading of GDPR, an organisation that self-hosts an Apache 2.0 model is the sole data controller. No third party has access to inference inputs. That is categorically different from API-based products where inputs are processed on vendor infrastructure.
VentureBeat reported that Thinking Machines Lab explicitly positions Inkling as a censorship-resistant model — meaning no external party controls how the deployed model behaves after you have it running.
Benchmark Performance
According to Artificial Analysis's evaluations, Inkling reaches an Elo rating of 1,238 on GDPval-AA v2 — an agent-based benchmark that simulates knowledge-work tasks — placing ahead of Kimi K2.6 (1,190) and DeepSeek V4 Flash (1,189) and leading US open-weight models in that index.
On technical benchmarks, Thinking Machines Lab reports:
- SWE-bench Verified: 77.6% — a coding benchmark measuring the ability to resolve real GitHub issues autonomously
- GPQA Diamond: 87.2% — a reasoning benchmark with graduate-level questions in physics, chemistry, and biology
For context: current local open-weight models in the 27B–32B class typically score in the 50–60% range on SWE-bench equivalent tasks, based on reported community measurements. Inkling operates at a noticeably higher quality tier, with the hardware requirements to match.
Deploying Today: APIs Without Big-Tech Intermediaries
Running the full Inkling model locally requires serious hardware — this is not a single Mac Studio deployment. For the majority of SMBs, local self-hosting of the full 975B model is a medium-to-long-term consideration tied to hardware availability and budget.
What is available today: Inkling is accessible via several API providers that have no connection to Google, Microsoft, or OpenAI — namely Together.ai, Fireworks.ai, Modal, Databricks, and Baseten. These providers run inference on their own infrastructure without forwarding requests for model improvement, which is a meaningful distinction from consumer AI products.
For SMBs that cannot build full on-premise AI infrastructure immediately, this is a practical path with substantially more data control than a standard SaaS subscription.
Inkling-Small: The Local Deployment Candidate
Alongside the flagship, Thinking Machines Lab is previewing Inkling-Small: a 276-billion-parameter model with 12 billion active parameters, trained on a similar recipe. The company states it matches or exceeds the larger model on several benchmarks, at significantly lower compute cost.
Inkling-Small is in preview at the time of writing — weights have not shipped publicly yet. When they do, the 12B active parameter footprint makes it a realistic candidate for local deployment on multi-GPU server configurations. For teams tracking data sovereignty strategy, this is the variant to watch.
What SMBs Should Do Now
Immediately:
- Test Inkling via Together.ai or Fireworks.ai — both offer API access without routing data through Big Tech infrastructure
- Download the Apache 2.0 weights from Hugging Face and build a development environment for internal experiments (requires a multi-GPU server or cloud GPU instance)
Near-term:
- Monitor Inkling-Small's release — when weights ship, the 12B active parameter footprint will likely be deployable on high-end local hardware without a full cluster
- Evaluate fine-tuning on internal data: Apache 2.0 means no IP complications when adapting the model to your documentation, contracts, or processes
For compliance-sensitive sectors:
- Assess which internal workflows — document analysis, email drafting, contract review, internal search — would benefit from frontier-quality reasoning, and whether current API access via Fireworks or Together meets your GDPR requirements. Based on our reading, that is considerably easier to justify than cloud SaaS products, provided proper data processing agreements are in place.
The Founding Story Matters
Thinking Machines Lab was founded by Mira Murati, who previously served as CTO of OpenAI. Choosing to release Inkling's full weights under Apache 2.0 — rather than a proprietary product or restricted community model — is a deliberate positioning: the company presents itself as an alternative to the Big Tech AI stack.
For European SMBs navigating EU AI Act deployer obligations and GDPR requirements, that distinction is more than a branding detail. It affects who controls the model you depend on, and what obligations follow from that dependency.
Next Steps
If you want to evaluate which local or private AI strategy fits your organisation — and how to deploy open-weight models like Inkling in a way that is technically sound and legally clean — we can help. Freshlab works with SMBs across the EU to build AI infrastructure that stays sovereign.
Start a pilot project — we respond within 24 hours.