not much happened today

2026-07-15 · Source: AINews · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

Thinking Machines Lab has launched Inkling, its first open-weights foundation model, an Apache 2.0 licensed multimodal Mixture-of-Experts (MoE) model. Inkling features 975B total parameters with 41B active parameters, trained from scratch on 45T multimodal tokens across text, image, and audio, supporting text output. It offers up to a 1M context window for open-weights checkpoints and 256K for its Tinker API. The release garnered broad day-0 ecosystem support from platforms like Hugging Face, vLLM, Modal, Baseten, and Databricks. Initial reception positions Inkling as the strongest U.S.-based open-weight model, scoring 41 on the Intelligence Index and #9 on the Agentic Web App Arena (Elo 1257), though it trails top Chinese open-weight and closed models on some benchmarks. Architectural innovations include relative positional encoding and scaled short convolution layers. An Inkling-Small variant (276B total / 12B active) is also noted for its competitive performance.

Key takeaway

For Machine Learning Engineers evaluating open-weight multimodal models, Inkling presents a compelling U.S.-based option. Its Apache 2.0 license, 1M context window, and broad ecosystem support make it highly adaptable for custom applications. You should consider Inkling for projects prioritizing fine-tuning, efficient reasoning, and integration into existing inference stacks like vLLM or Modal, rather than solely chasing benchmark supremacy. Explore Inkling-Small for optimized cost and latency.

Key insights

Inkling's launch emphasizes an open-weight, customizable multimodal foundation model designed for practical application and ecosystem integration over pure benchmark leadership.

Principles

• Open-weight models with permissive licenses drive ecosystem adoption and customization.
• Prioritize a tunable foundation model over benchmark-maxing for strategic differentiation.
• Architectural experimentation, like relative positional encoding, can signal future trends.

Method

Inkling employs a Mixture-of-Experts architecture with 2 shared expert sinks, relative positional encoding, and scaled short convolution layers. Inference is optimized with sconv-aware tensor-parallel sharding and DFlash speculative decoding.

In practice

• Deploy Inkling via vLLM, Modal, or Hugging Face for multimodal capabilities.
• Apply 1-bit GGUF quantization to reduce model footprint by 86% for local inference.
• Integrate DFlash speculators or FlashAttention-4 for up to 67% throughput gains.

Topics

• Inkling
• Mixture-of-Experts
• Multimodal AI
• Open-Weight Models
• Model Inference
• AI Quantization

Code references

• PrismML-Eng/llama.cpp
• PrismML-Eng/mlx
• z-lab/dflash
• PrismML-Eng/Bonsai-demo
• soofi-project/Soofi-Pretraining

Best for: CTO, VP of Engineering/Data, AI Architect, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• [AINews] Thinky's Inkling: 975B-A41B multimodal, new best American Apache 2.0 open model (with Inkling-Small, 276B-A12B) — Inkling is an Apache 2.0 multimodal MoE model, prioritizing customization and efficiency over raw benchmark supremacy.
• Thinking Machines Lab Drops Its First Model — Thinking Machines Lab launched Inkling, a large open-source multimodal model, challenging established AI players with a vision for decentralized AI.
• Together AI brings Thinking Machines Lab’s new model Inkling on day 0 — Inkling unifies multimodal inputs and offers controllable reasoning via a novel MoE architecture for broad task versatility.
• The AI That Can Re-Write Its Own Brain: Why Inkling is the New Frontier for Open Weights — Inkling enables autonomous AI development and dynamic intelligence scaling through an open-weights, self-modifying architecture.
• Inkling: Our open-weights model — Inkling is an Apache-2.0 licensed multimodal base model for fine-tuning, not a frontier model.
• Inkling model from Thinking Machines Lab now on Databricks — Inkling, an open-weights coding model, is now on Databricks, offering enterprises customizable, secure, and cost-effective AI for coding workflows.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AINews.