Are the Inkling 2 Small weights open?

Not yet. Inkling 2 Small is available through Tinker with closed weights. Thinking Machines Lab says it intends to release the full weights under Apache 2.0 next week, after external safety testing and red-teaming. That is a stated plan, not a confirmed date.

How does Inkling 2 Small compare with the first Inkling on safety?

Thinking Machines reports an internal misbehavior rate of 14.9 for Inkling 2 Small, against 66.8 for the first Inkling (lower is better). The post also says the model still trails GPT-6.1 Sol on its alignment evaluations.

How does Inkling 2 Small do on coding benchmarks?

Thinking Machines lists 72.1% on DeepSWE v1.1, 86.3% on SWE-Bench Pro v2 (Hard) and 81.6% on Vibe Code Bench v1.1 for Inkling 2 Small. The post compares these with about a dozen open and closed models in its own tables.

Does Inkling 2 Small support audio or images?

Inkling 2 Small accepts text and image inputs and produces text. It does not support audio input, unlike the earlier Inkling-Small. The vision encoder has 292M parameters and is based on Perception Encoder.

Thinking Machines Lab · 2026-11-05 · major

Inkling 2 Small — Thinking Machines' 594B MoE, weights due next week

Inkling 2 Small is a 594B-parameter MoE model with 17B active, now on Tinker with closed weights. Thinking Machines says Apache 2.0 weights follow next week, after safety testing.

Cover graphic for the Inkling 2 Small announcement from Thinking Machines Lab

Thinking Machines' second-generation model is a 594B MoE with 17B active parameters, live on Tinker.

Quick facts

Maker	Thinking Machines Lab
Parameters	594B total, 17B active (MoE)
Context window	Up to 1M tokens
Inputs	Text and image (no audio)
Availability	Tinker, closed weights for now
License	Apache 2.0 planned for weights
AA Intelligence Index	31

What is it?

Inkling 2 Small is the first model of Thinking Machines Lab's second Inkling generation. It is a sparse mixture-of-experts transformer with 594B total and 17B active parameters. It reads text and images, supports variable reasoning effort, and handles context up to 1M tokens.

How does it work?

The model was pretrained natively in NVFP4 on NVIDIA GB300 NVL72 racks, 8.4T tokens in the first phase. The main reinforcement-learning run took 1,075 steps and over 17M rollouts across 9 days on 2,048 GPUs. Long tasks use compaction, so agents can work past the 256K training context.

Why does it matter?

Thinking Machines reports a score of 31 on the Artificial Analysis Intelligence Index. For teams that want a US open-weight base to fine-tune, the Apache 2.0 release planned for next week is the part to watch.

Who is it for?

teams fine-tuning open models, coding-agent builders

Frequently asked questions

Are the Inkling 2 Small weights open?: Not yet. Inkling 2 Small is available through Tinker with closed weights. Thinking Machines Lab says it intends to release the full weights under Apache 2.0 next week, after external safety testing and red-teaming. That is a stated plan, not a confirmed date.
How does Inkling 2 Small compare with the first Inkling on safety?: Thinking Machines reports an internal misbehavior rate of 14.9 for Inkling 2 Small, against 66.8 for the first Inkling (lower is better). The post also says the model still trails GPT-6.1 Sol on its alignment evaluations.
How does Inkling 2 Small do on coding benchmarks?: Thinking Machines lists 72.1% on DeepSWE v1.1, 86.3% on SWE-Bench Pro v2 (Hard) and 81.6% on Vibe Code Bench v1.1 for Inkling 2 Small. The post compares these with about a dozen open and closed models in its own tables.
Does Inkling 2 Small support audio or images?: Inkling 2 Small accepts text and image inputs and produces text. It does not support audio input, unlike the earlier Inkling-Small. The vision encoder has 292M parameters and is based on Perception Encoder.

Try it

Tinker Playground: https://tinker.thinkingmachines.ai/playground