Inkling

Open weights 975B multimodal model built for fine-tuning

Artificial IntelligenceDevelopment

▲ 165 votes11 commentsLaunched Jul 20, 2026

Visit Website

Daily #7Weekly #6

Inkling is Thinking Machines’ first open-weights model, a 975B MoE with 41B active parameters, 1M context, native reasoning across text, images, and audio, and controllable thinking effort. Fine-tune it on Tinker or download the Apache 2.0 weights.

AI Analysis

📝 Summary

Inkling is Thinking Machines’ first open-weights 975B MoE model with 41B active parameters, featuring 1M context, native multimodal reasoning (text, images, audio), and controllable thinking effort. Core features include fine-tuning via Tinker platform or direct Apache 2.0 weights download. It solves pain points of limited customization, high training costs, and lack of open multimodal models for developers and researchers. USP is efficient, adaptable high-performance AI that democratizes advanced capabilities for specialized applications without building from scratch. Value proposition centers on flexibility, transparency, and community-driven innovation in AI development.

📈 Market Timing

In 2025-2026, explosive growth in open-source AI, maturing MoE and multimodal tech, rising demand for customizable models due to data privacy regulations and cost pressures make it highly favorable. Industry shift from closed APIs to open weights aligns perfectly with user needs for control and specialization. Excellent Timing.

✅ Feasibility

High technical difficulty for such a large model, but MoE design enables efficient inference. Substantial development costs already incurred; operational costs for Tinker platform are notable but scalable via community. Moderate compliance risks (AI ethics, licensing). Strong scalability potential as open weights. Overall rating: High.

🎯 Target Market

Main segments: AI/ML engineers, researchers, tech startups and enterprises in software dev, content/media, and multimodal AI apps. Primarily US, Europe, Asia tech hubs. TAM for open AI models/tools ~$20B+ by 2026; SAM for multimodal fine-tuning ~$2-5B; SOM for this scale smaller but growing. Pain points: barriers to custom multimodal AI and compute costs. High willingness to pay for fine-tuning services, enterprise support.

⚔️ Competition

Competition level: Medium. Direct competitors: 1. Mixtral (mistral.ai), 2. Llama 3 (meta.com/llama), 3. DeepSeek-V2 (deepseek.com), 4. Qwen2-VL (qwen.ai), 5. DALL·E 3 / GPT-4o open alternatives (openai.com but less open). Advantages: larger scale with efficiency, 1M context, full native audio support, controllable effort, permissive Apache license. Disadvantages: newer with less established ecosystem and potentially higher resource needs vs. optimized smaller models.

Upgrade Pro to unlock full AI analysis