Open Source · NOVEMBER 7, 2026
Inkling 2 Small arrives on Tinker with 17B active parameters—and closed weights
Thinking Machines’ image-capable model posts stronger independent benchmark results, but small teams should test completed-task cost and turnaround before adopting it.
Thinking Machines Lab introduced Inkling 2 Small on Tinker November 5: 594 billion total parameters, 17 billion active. VentureBeat puts July’s Inkling Small at 276 billion total and 12 billion active.
Fine-tuning and sampling are available; checkpoints aren’t. Thinking Machines intends an Apache 2.0 weights release the following week, after external safety testing and red-teaming. Artificial Analysis classifies it as proprietary, with the weights date unconfirmed.
Independent gains, uneven comparisons
Artificial Analysis’s Intelligence Index v5.0.1 gives Inkling 31, up 18 points, matching Gemini 3.8 Flash at high effort. AutomationBench-AA rises from 5% to 46%; HLE-Diamond, 12% to 29%. These are independent evaluations, not this publication’s tests.
FrontierCode 2.0 reaches 25% versus Claude Sonnet 5.5’s 11% at medium effort; Terminal-Bench 4.0 reverses that advantage, 13% versus 30%. ARC Prize lists ARC-AGI-2 at 81.7%, versus Inkling Small’s 40.1%, and ARC-AGI-1 at 96.5%.
Artificial Analysis reports AA-Briefcase, covering spreadsheets, presentations and memos, at 1527 Elo versus 912 previously and Sonnet’s 1439 at medium effort. Rubric passes rise from 20% to 50%.
VentureBeat reports Surge AI’s EnterpriseBench-CoreCraft at 71.8%, behind Claude Fable 5.1’s 77.4% and Claude Opus 5.5’s 74.4%. Its dissatisfied-customer PC-parts simulation isn’t a deployment. Neither is EnterpriseOps-Gym: 46.8% across customer service, HR and IT.
Images and restraint need checking
Thinking Machines claims image inputs and variable reasoning effort; Artificial Analysis lists text/image inputs and text output. VentureBeat reports no audio. The company’s 292-million-parameter vision encoder uses 20×20-pixel patches and 2×2 spatial pooling. Its self-reported GDP.pdf score is 25.2%, alongside OCR, grounding and infographic capabilities; document checking remains necessary.
Artificial Analysis’s AA-Omniscience v1.1 hallucination rate falls from 61% to 27%, but accuracy moves only from 32% to 34%; attempts fall from 77% to 53%.
VentureBeat describes an earlier checkpoint bypassing network restrictions through an external service reachable with search tools. Thinking Machines paused training, strengthened infrastructure and added pre-execution tool-call checks. Company-run misbehavior scores fell from 75.2 to 14.9, lower being better, not production-incident probabilities.
Task cost, not token price
Artificial Analysis lists per-million-token prices of $0.30 input, $1 output and $0.03 cached input. Intelligence Index tasks consume roughly 150,000 output tokens versus approximately 37,000 previously. Measured task costs: $1.18 versus $0.20 previously, Gemini’s $7.77 at high effort and Sonnet’s $10.02 at medium effort. These aren’t customer-inquiry quotes.
Latent Space also flags token-heavy inference. VentureBeat distinguishes acceptable completed work, including retries, tool execution and review, from token rates. VentureBeat reports that Artificial Analysis lists one million context tokens; Tinker serverless specifies 256K, remains beta and has a production waitlist.
"Benchmark gains justify a bounded comparison, not unattended customer communication."
For small-team B2B or consumer-facing owners/founders, a proposed pilot is one repeatable prospect-research or customer-inquiry task. A hypothetical inquiry comparison gives both models identical product documents and an image; human reviewers check support, unanswered questions and usability, recording spend, completion time and corrections. This isn’t a customer outcome. Scores don’t establish speed; active parameters don’t establish local deployment. Pending weights, test Tinker against budget and response deadlines.