0
Neue Alchemy
Sign inSign upLet’s talk
Signal + NoiseIntelligence, distilled daily.
MODEL SIGNAL · THINKING MACHINES · NEW

Inkling Small

An open-weights, 276-billion-parameter sparse Mixture-of-Experts multimodal transformer (276B total / 12B active) designed to deliver much of the flagship Inkling model’s reasoning, coding, and multimodal performance at lower compute cost and latency.

CATEGORYGeneral
CONTEXT1000000
RELEASEDJuly 30, 2026
Key Features
  • Sparse Mixture-of-Experts (MoE) multimodal decoder-only transformer
  • 276B total parameters with 12B active per token
  • Open-weights release under the Apache 2.0 license
  • Accepts text, image, and audio inputs and generates text outputs
  • Supports a context window of up to 1M tokens
  • Designed to match or exceed the larger Inkling model on many published benchmarks while offering lower cost and latency

Provider announcement →

Read the Model Signal report →

MODEL SIGNAL

Inkling Small

Thinking Machines brings a massive 1M-token context and true multimodal inputs to the open-weights arena with an Apache 2.0-licensed MoE.

Bottom line

Released on July 30, 2026, Thinking Machines' Inkling Small enters the ecosystem as a 276-billion-parameter Sparse Mixture-of-Experts (MoE) transformer. Despite the "Small" moniker, it brings flagship-grade specs to self-hosted pipelines: an Apache 2.0 open-weights license, a 1-million-token context window, and multimodal input handling for text, images, and audio.

Signal

The confirmed facts from the provider's model card and release announcements establish a highly capable architecture. Inkling Small operates on a massive 276B total parameter base but routes to just 12B active parameters per token. This decoder-only transformer natively accepts text, image, and audio inputs to generate text outputs. The most definitive signal for operators is the Apache 2.0 license combined with a 1,000,000-token context window, placing it in a rare class of highly permissive, extremely long-context multimodal open-weights models.

Noise

Thinking Machines explicitly designed this release to match or exceed their larger flagship Inkling model on published benchmarks while driving down cost and latency. As is standard at launch, these provider-stated benchmark equivalencies remain directional noise until third-party operator validation scales. Additionally, the term "Small" is highly relative; while the 12B active parameters suggest fast inference compute, the 276B total parameter count means the static memory footprint remains substantial.

Operator implications

The directional read for operators is that they get access to massive routing capacity without paying the continuous compute tax of a dense 200B+ model. However, operators must separate compute speed from VRAM requirements. Loading a 276B open-weights model requires significant GPU memory, meaning this is not a single-GPU workstation model. It is built for multi-node enterprise deployments where permissive licensing and data privacy are non-negotiable but hardware budgets are robust.

Where it fits

Inkling Small fits squarely in enterprise pipelines that require massive context ingestion—such as full codebase analysis, lengthy document summarization, and mixed-media processing—where data cannot be sent to closed-API providers. The native audio and image ingestion combined with the 1M window makes it a strong candidate for complex multimodal RAG architectures. It does not fit edge deployments or memory-constrained hosting environments.

Model Signal · Signal + Noise · Isaiah Steinfeld