Thinking Machines Open-Sources Inkling

Latent Space · rss · 2026-07-16

Latent Space summarized the first open-weight foundation model family, Inkling, released by Thinking Machines Lab.

Core Specs

• Inkling: MoE Transformer, approx. 975B total parameters / 41B active parameters
• Supports text, image, audio, with text output
• Pre-trained on approx. 45T tokens, with context lengths up to 1M
• Licensed under Apache 2.0
• A smaller variant, Inkling-Small (276B / 12B), is also available, focusing on lower cost and latency

Training & Architecture Highlights

• Trained from scratch, emphasizing controllable thinking/reasoning effort
• The community uncovered several architectural details from the materials:
• Mixed/sliding window attention
• Relative position encoding/relative attention bias
• Short convolution layers
• MoE + shared experts
• Load balancing scheme
• Multi-head design for speculative decoding

Performance & Positioning

• Artificial Analysis gave it an Intelligence Index of 41, calling it one of the leading US open-weight models currently available
• It also achieved strong rankings on leaderboards like the agentic / web app arena
• Multiple reviewers noted its solid performance in tool calling, long-horizon error correction, and concise reasoning, though it may not fully surpass top closed-source models across all multimodal and agentic benchmarks

Ecosystem

• Received ecosystem support from vLLM, SGLang, Modal, Baseten, Databricks, and Hugging Face on launch day
• Tinker platform and Playground are also available, supporting subsequent fine-tuning and usage

Original post →

More from Models

• Developer Building a Unified Leaderboard of All Model Benchmark Scores — airesearch12 · 2026-09-11
• Rumor claims Kimi faked performance by serving Claude; DeepSeek new model surprises in evals — realsohamparekh · 2026-09-11
• GPT-5.6 writes well but is instantly forgettable, user complains — BasedRaddka · 2026-09-11
• Opus Refuses Protein Research Codebase Over 'Safety' Concerns, Dev Considers Rolling His Own — josephdviviano · 2026-09-11
• User Hails Unconfirmed 'DeepSeek 4.1 Flash' as an Inflection Point in LLMs — himanshustwts · 2026-09-11
• Terminal Bench v4: GLM-5.3 Leads at 41.9%, Kimi-K3 Underwhelms at 12.6% — Ok_Warning2146 · 2026-09-11

Models channel →