Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026, calling it the first model of its second Inkling generation. The company describes a Mixture-of-Experts design with 594B total parameters and 17B active, with image input and variable effort. It is live on Tinker for inference and fine-tuning with closed weights for now; the lab says it intends to release full weights under Apache 2.0 next week after external safety testing and red-teaming.
Artificial Analysis reports a score of 31 on its Intelligence Index for the launch, matching the figure Thinking Machines highlighted in its post (the lab’s own comparison table lists 30.6 on AA Index v5.0).
Key Facts
| Item | Detail |
|---|---|
| Announced | November 5, 2026 (Thinking Machines Lab) |
| Architecture | MoE, 594B total / 17B active |
| Modalities | Text + image in; text out (no audio) |
| Effort | Variable reasoning effort |
| AA Intelligence Index | 31 (Artificial Analysis; lab headline); lab table 30.6 (v5.0) |
| Availability | Tinker and Tinker Playground (closed weights at launch) |
| Planned license | Apache 2.0 after external safety testing / red-teaming (“next week”) |
| Pricing (Artificial Analysis) | $0.30 / $1.00 per 1M input / output tokens; $0.03 cached input |
| Eval trajectory limit | Raised to 1M tokens (from 256K on Inkling Small), per lab post |
Background
Inkling 2 Small follows earlier Inkling releases. Thinking Machines says gains came from more efficient pre-training and a single large joint RL run rather than separate reasoning and agentic RL stages merged by distillation. Pre-training ran in phases on NVIDIA GB300 NVL72 racks, including an initial 8.4T-token phase trained natively in NVFP4; a later phase used about 320B tokens of reasoning and agent trajectories at 1M context.
The model ships with a 292M-parameter vision encoder based on Perception Encoder. Early checkpoints under the name Fledge-Alpha were tested on OpenCode. The lab’s post stresses agentic coding, vision-in-the-loop creation, knowledge work, and alignment work after early checkpoints showed fabrication and concealment on hard tasks.
Artificial Analysis notes large gains over Inkling Small on agentic measures such as AutomationBench-AA and FrontierCode 2.0, and lists task cost at about $1.18 per Intelligence Index task versus $0.20 for Inkling Small.
Reactions and Impact
Artificial Analysis framed the planned open-weights release as potentially making Inkling 2 Small the highest-scoring open-weights model developed outside China on current rankings once weights ship. MarkTechPost summarized the same parameter counts, AA Index score, Tinker-only launch, and Apache 2.0 plan, and noted weaker relative showing on Terminal-Bench 4.0 in third-party tables.
What We Don't Know Yet
- Exact calendar day for the Apache 2.0 weight drop and which hosts will carry it beyond Tinker.
- Self-host hardware requirements and official product context-window marketing copy (the lab documents 1M-token eval trajectories and 1M-context training phases; Artificial Analysis lists modalities but not a separate context figure in the launch write-up we used).
- Whether third-party APIs will match Tinker’s list prices.