Former OpenAI CTO Mira Murati's startup Thinking Machines Lab released Inkling, a 975B-parameter open-weights MoE model, per a recent Grok AI News thread linked to the WSJ report.
The release targets developers seeking alternatives to closed frontier systems.
Model: Inkling | Parameters: 975B MoE | License: Open-weights | Focus: Customizability
What Inkling Is and How It Works
Inkling follows architectures from leading Chinese open-source models. It uses a mixture-of-experts design that activates only subsets of its parameters during inference.
The model ships with open weights, allowing full fine-tuning and modification. This approach differs from API-only releases by OpenAI and Anthropic.
Key Specs and Architecture Details
The 975B total parameters activate fewer experts per token, reducing compute needs compared with dense models of similar scale. No public benchmark scores were released with the launch.
The design prioritizes customizability over immediate out-of-the-box performance on standard leaderboards.
How to Try Inkling
Weights are available through the company's initial distribution channels. Developers can download and run the model on clusters supporting MoE inference frameworks.
No hosted API was announced at launch. Users must handle their own deployment and fine-tuning pipelines.
Tradeoffs of the Open-Weights Approach
- Full weight access enables domain-specific adaptation unavailable in closed models.
- Requires significant hardware for full-scale inference.
- No safety guardrails are enforced by the provider after download.
- Early community testing shows variable instruction-following compared with GPT-4-class systems.
Competing Open Models
Inkling enters a field already populated by large open releases.
| Feature | Inkling | Qwen2.5-72B | Llama 3.1 405B |
|---|---|---|---|
| Size | 975B MoE | 72B dense | 405B dense |
| Weights | Open | Open | Open |
| Focus | Customizability | General | General |
| Origin | US startup | Chinese | US lab |
Who Should Use Inkling
Research teams and companies needing full model control will find the open weights useful. Organizations already running large MoE workloads can integrate it without new infrastructure.
Teams requiring immediate high benchmark scores or managed safety layers should continue with closed APIs instead.
Verdict on the Release
Inkling marks the first concrete step by Thinking Machines Lab toward providing modifiable frontier-scale models outside the major labs.
Bottom line: The 975B MoE release gives practitioners direct weight access at a scale previously limited to closed providers.
The move tests whether open weights at this size can sustain independent development ecosystems.
Top comments (0)