Thinking Machines Releases Inkling Small, a 12B-Active-Parameter Model That Beats Its Larger Predecessor on Key Benchmar
Thinking Machines has released Inkling Small, an open-weights reasoning model with 276 billion total parameters but only 12 billion active. According to Artificial Analysis, it scores nearly as high as the company's larger Inkling model while using roughly a third of the parameters and far fewer output tokens per task.
Thinking Machines, the AI lab founded by former OpenAI CTO Mira Murati, has released its second model, Inkling Small, prioritizing efficiency over raw scale.
According to Artificial Analysis, the open-weights reasoning model scores 40 on the Intelligence Index — just one point below the company's first model, Inkling, which scores 41. Inkling Small achieves this with 276 billion total parameters and only 12 billion active parameters, less than a third of what its larger sibling uses. Artificial Analysis states that no open model of equal or smaller size scores higher on its Intelligence Index.
Benchmark performance
Inkling Small outperforms Inkling on several coding and reasoning benchmarks despite its smaller active parameter count:
- Humanity's Last Exam: 32% vs. 30% for Inkling
- GPQA Diamond: 89% vs. 87% for Inkling
The smaller model falls behind on agent-based tasks and factual knowledge benchmarks, according to Artificial Analysis, suggesting a tradeoff between reasoning efficiency and broader knowledge retention.
Token efficiency
The standout metric is token efficiency. Inkling Small averages 24,000 output tokens per task, compared to 45,000 for DeepSeek V4 Flash and 78,000 for GPT-5.4 mini, according to Artificial Analysis data. Lower token consumption per task translates directly into lower inference costs and faster response times for equivalent reasoning quality — a significant factor for production deployments running at scale.
Specifications
Inkling Small handles text, image, and speech inputs and ships with a 256,000-token context window. Thinking Machines has released the model under the Apache 2.0 license, with weights available on Hugging Face. Pricing for API access has not yet been disclosed.
Users can fine-tune the model directly in the browser through Thinking Machines' Tinker Playground. The company positions Inkling Small — like its predecessor — as a foundation model meant to be fine-tuned with customer-specific data rather than used purely out of the box.
What this means
Inkling Small is a bet that mixture-of-experts efficiency, not parameter count, will define competitive open models going forward. A model with 12 billion active parameters matching or beating a larger sibling on reasoning benchmarks — while using roughly a third the output tokens of comparable models from DeepSeek and OpenAI — has real cost implications for anyone running reasoning workloads at volume.
The agentic task and factual knowledge gaps are worth watching. If Inkling Small underperforms on tasks requiring broad world knowledge or multi-step tool use, it may be best suited for narrow reasoning-heavy applications rather than general-purpose agents. Thinking Machines' emphasis on fine-tunability via Tinker Playground also signals a strategy: rather than competing purely on raw capability, the company is betting that customizability on top of an efficient base model is what enterprise customers actually want. Whether that bet pays off depends on adoption data that isn't yet available — Artificial Analysis benchmarks measure raw capability, not real-world fine-tuning outcomes.
Related Articles
Thinking Machines Lab Releases Inkling Small: 276B MoE Model with 524K Context Window
Thinking Machines Lab has released Inkling Small, an open-weight multimodal mixture-of-experts model with 12B active parameters out of 276B total and a 524K token context window. The model targets reasoning, coding, agentic workflows, and multilingual use cases at $0.58 per 1M input tokens and $1.44 per 1M output tokens.
AMD Releases Instella-MoE-16B-A3B-Think, a Fully Open Mixture-of-Experts Model Trained Entirely on AMD GPUs
AMD has released Instella-MoE-16B-A3B-Think, a 16-billion-parameter Mixture-of-Experts language model trained entirely from scratch on AMD Instinct MI300X and MI325X GPUs. The release includes every checkpoint from pre-training through reinforcement learning, along with full training recipes, under a research-only license.
DeepSeek Releases V4 Flash 0731: 1M-Token MoE Model at $0.14/M Input Tokens
DeepSeek has released V4 Flash 0731, a sparse mixture-of-experts model with 13B active parameters out of 284B total and a 1049K token context window. The model targets coding, reasoning, and agent workflows, priced at $0.14 per million input tokens and $0.28 per million output tokens via OpenRouter.
Unsloth Releases GGUF Quantizations of Kimi K3, a 2.8T-Parameter Open-Weight MoE Model
Unsloth has released GGUF quantizations of Kimi K3, a 2.8-trillion-parameter open-weight Mixture-of-Experts model from Moonshot AI with a 1-million-token context window and native vision support. The largest lossless quantization (Q8) weighs in at 1.56TB.
Comments
Loading...