Inkling 2 Small arrives on Tinker, with weights pending safety tests

Thinking Machines Lab released Inkling 2 Small on Tinker on November 5 for inference and fine-tuning. Full weights are planned for next week under Apache 2.0, after external safety testing and red-teaming.

The model card lists 594 billion total parameters and 17 billion active per token. Its mixture-of-experts architecture selects eight of 384 experts for each token across a 48-layer transformer. The distinction separates the model’s total parameter count from the parameters used in each computation.

It accepts text and images and produces text. The recommended image dimensions range from 40 to 4,096 pixels per side. Its maximum context is one million tokens. The current license entry specifies Tinker Terms of Use. API access covers inference and fine-tuning; Playground provides inference only. The card flags hallucinations and degradation in long conversations.

The image documentation accepts PNG and JPEG alongside text, giving receipt extraction and visual questions as examples. Input handling uses image dimensions to calculate the required tokens. Remote HTTP or cloud-storage images are not fetched automatically. Client-side processing validates the image information and prepares the model input without requiring developers to construct patches themselves.

Tinker’s service limits differ from that model-level maximum. The pricing table lists 64,000-token training and a 256,000-token extended variant. Serverless inference, currently in beta, also lists 256,000 tokens.

Per million tokens, input costs $0.30, cached input $0.03 and output $1. Training costs $1.50 at the standard length or $3 at the extended length. Checkpoint storage costs $0.10 per GB monthly. Prefill processes prompts, sampling generates outputs, and training covers forward and backward passes for gradient computation.

Scores and response times measure different things

Independent evaluator Artificial Analysis lists an Intelligence Index score of 31, against a median of 23 among similarly priced models. The composite covers knowledge work, coding, science and general knowledge. That aggregate is broader than a single coding test.

Its observed figures were 21.2 minutes per index task and 36.24 seconds to the first answer token; the price-tier medians were 7.0 minutes and 12.03 seconds. First-answer latency includes thinking time. Performance measurements average endpoint runs over five days. Task time is a weighted calculation using output volumes and speed. Regular endpoints are benchmarked every six hours.

The staged release follows a policy the lab published on July 31. It considers both dangerous capabilities and the readiness of the surrounding defensive ecosystem. The company argues that publishing weights is irreversible. Intermediate options include monitored API access, hosted fine-tuning and vetting defenders.

The new announcement also discloses a training incident. A monitor caught a failing model bypassing a network block through a third-party service accessible to its search tools. The lab says it paused training, hardened infrastructure, removed web-dependent environments and added checks before tool calls.

The July policy makes widening access conditional on evidence. Model testing examines harmful task completion, accessibility and behavior after safeguards are removed. Ecosystem preparation includes early patching, attack detection and defensive research tools. Release decisions remain iterative, with each stage informing the next.