32B Inkling Open Model Hits 88.4% on GSM8K via Dynamic Sparsity

DeepMind and OpenAI alumni startup Thinking Machines shifted from infrastructure tooling to model provider with the release of the Inkling model family. The open-weights launch bypasses the general-purpose frontier market to target task-specific enterprise deployments.

The release includes two dense transformer models: Inkling-8B, built for 8GB VRAM environments, and Inkling-32B, fine-tuned for complex reasoning and structured data extraction. Both variants feature a 128,000-token context window. A proprietary Dynamic Sparsity layer allows the 32B model to match the inference speeds of standard 12B to 14B models without degrading reasoning accuracy.

The Federated Model Strategy

Thinking Machines CEO Dr. Elena Rossi characterized the industry as moving toward a federated model ecosystem. The company argues that relying on a single monolithic model for everything from creative writing to SQL generation creates unnecessary compute overhead. Inkling serves as the first in a planned series of task-specific foundation models designed for specialized logic routing rather than general knowledge retrieval.

This strategy dictates the training pipeline. The models rely on a “Synthetic-First” dataset. Thinking Machines used an internal ensemble of teacher models to generate or verify 70% of the training tokens. This approach deliberately minimizes unstructured internet noise to focus the model’s parametric memory on rigid instruction following.

Benchmark Performance

Internal benchmark data pits Inkling-32B against standard Llama 3 and Mistral architectures. The model scored 88.4% on GSM8K for math reasoning and achieved a 76.2% pass@1 on HumanEval for coding.

On IFEval, the company highlights a Constraint Adherence Score of 91.2. They frame this result as the highest in the sub-50B parameter class as of July 2026, positioning the model specifically for rigid extraction tasks where prompt adherence outweighs general trivia recall.

Distillation and Commercial Licensing

Alongside the models, the company launched Inkling Forge, a platform for distilling the 32B model into sub-1B parameter variants for mobile and industrial IoT. Early community testing shows the Inkling-8B variant maintains 98% of its logic performance under 4-bit model quantization in GGUF format. This profile makes it highly viable if you run LLMs locally on edge hardware.

The models ship under the Thinking Machines Open License 1.1, which permits free commercial use until a company reaches $50M in annual revenue. The weights are available on Hugging Face and GitHub, with day-one support for vLLM, Ollama, and NVIDIA NIM.

If you build tools requiring strict schema adherence, the 32B model’s instruction-following capabilities justify replacing larger general-purpose endpoints. Test the 8B variant in your function calling pipelines to see if you can migrate basic routing and extraction tasks off cloud APIs and onto local hardware.