A Guth Labs publication ↗
Back to News

Inkling 2 Small moves from hosted access to downloadable weights

Inkling 2 Small moves from hosted access to downloadable weights

Thinking Machines Lab has released downloadable weights for Inkling 2 Small, changing the access story from its November5 hosted introduction. A clearly dated November12 update on the original announcement says the weights are available on Hugging Face; the official repository now displays Apache 2.0 licensing. Dated announcement · Official repository

The model is a mixture of experts with 594 billion total parameters and 17 billion active, accepts images and supports variable reasoning effort. Thinking Machines lists serving integrations including vLLM and SGLang, and API availability through providers including Cloudflare. These are the lab’s stated integrations, not hands-on deployment tests.

The page still retains the original hosted-only language below its update. Readers should not mistake that historical launch paragraph for the current availability status. November12 is the weights-release update; November5 remains the original introduction date.

Analysis: open access changes the options, not the hardware math

The consequential development is control. Downloadable weights let capable operators evaluate a model outside the original hosted service and investigate customization or their own serving arrangements. That is a different proposition from a promise to publish weights later.

But “Small” is a family label, not a laptop recommendation. Active parameters indicate which parts participate in computation; the much larger total still matters for storage and deployment. An operator should establish memory, quantization, context and concurrency requirements for a specific setup before choosing hardware. This article does not give a tested minimum configuration.

Provider and hardware update, November 15, 2026

Publication review integrates the November 13 provider addenda; the weights-release event remains November 12. Cloudflare documentation lists model ID @cf/thinkingmachines/inkling-2-small, text/image input, reasoning, tool calling and a 1,048,576-token context window. Listed prices per million tokens are $0.30 input, $1.00 output and $0.10 cached input, with Workers Paid or prepaid AI Gateway credits required. These mutable rates are a review-time snapshot, not a quote.

Cloudflare describes 18B active parameters, while the lab model card says 17B. This article retains the lab-attributed figure rather than silently resolving that discrepancy. The lab specifies at least 500GB aggregate GPU memory for NVFP4 or 800GB for MXFP8. These are vendor requirements, not Guth-tested minima.

Ollama’s listing offers the model with a :cloud command. Its posted input/cached/output rates are $0.30/$0.03/$1.00 per million tokens. Pay-as-you-go has no subscription fee but charges for inference. This is hosted access, not evidence of laptop execution.

Ollama’s March 2026 privacy policy says cloud prompts and responses are processed transiently, not trained on; account and usage metadata are separate. These are provider statements, not an independent privacy audit.

Analysis: Familiar client software can hide a changed execution boundary. Record whether a call ran on local hardware or a hosted provider, and evaluate cost, latency and data handling separately.

Evidence limits: Primary model documentation and provider pages were directly read; the canonical Cloudflare URL initially failed but its indexed version loaded. No inference, download or independent performance test was performed. Exact provider launch times are unverified.

Sources and citations

Revision history

Published .