What Is Inkling 2 Small? Tinker Pricing and Closed Weights (2026)
Contains affiliate links; your price doesn't change · Disclosure · How we research
TL;DR
Inkling 2 Small is Thinking Machines Lab's second-generation model, released November 5, 2026. It is a 48-layer decoder-only MoE with 594B total parameters and 17B active, text plus image input, and a model-card context window up to 1M tokens. It ships on Tinker with closed weights under Tinker Terms of Use. As of November 9, 2026 Taipei, full Apache 2.0 weights are still pending. On Tinker, 64K and 256K training SKUs share prefill $0.30 / cached prefill $0.03 / sample $1.00 per 1M tokens; train is $1.50 (64K) or $3.00 (256K). Serverless inference Beta lists 256K at input $0.30 / cached $0.03 / output $1.00 per 1M.
How this article was made: This article is research-based. Launch claims come from Thinking Machines Lab's blog Introducing Inkling 2 Small (November 5, 2026). Architecture, modalities, license, and context come from the official model card. Token rates and Tinker IDs come from the Tinker pricing page and Tinker Models & Pricing docs. The AA Intelligence Index score of 31 is vendor-reported, not our benchmark. We did not invent VRAM tables or claim weights are public. Checked again November 9, 2026 (Taipei time). Display name: CommX.
Inkling 2 Small is Thinking Machines Lab's November 5, 2026 second-generation model. It is a MoE with 594B total / 17B active parameters, text plus image input, and closed weights on Tinker today.
If you only need the buy path: use Tinker for fine-tuning and sampling now; treat Apache 2.0 full weights as still pending as of this check.
What is Inkling 2 Small?
Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026. Official materials describe:
- Architecture: 48-layer decoder-only MoE; each token routed to 8 of 384 experts
- Size: 594B total parameters, 17B active
- Numerics: MXFP8 and NVFP4
- Inputs: text and image (image dims ideally 40px–4096px); no audio (relative to prior Inkling)
- Context (model card): up to 1M tokens
- Distribution: Tinker for fine-tuning and inference; playground available
- License today: Tinker Terms of Use (closed weights)
The blog also cites an Artificial Analysis Intelligence Index score of 31 (vendor-reported).
Inkling 2 Small announcement (official)
How much does Inkling 2 Small cost on Tinker?
Rates below are from the official Tinker pricing page and Models & Pricing docs. Checked November 9, 2026 (Taipei). Currency is USD. Thinking Machines / Tinker is not in our affiliate list; links are plain official URLs.
Training & sampling (USD per 1M tokens):
| SKU | Tinker ID | Prefill | Cached prefill | Sample | Train |
|---|---|---|---|---|---|
| Inkling 2 Small 64K | thinkingmachines/Inkling-2-Small |
$0.30 | $0.03 | $1.00 | $1.50 |
| Inkling 2 Small 256K | thinkingmachines/Inkling-2-Small:peft:262144 |
$0.30 | $0.03 | $1.00 | $3.00 |
Serverless inference Beta (Inkling family only; USD per 1M):
| SKU | Context | Input | Cached input | Output |
|---|---|---|---|---|
| Inkling 2 Small | 256K | $0.30 | $0.03 | $1.00 |
Checkpoint storage: $0.10 / GB / month.
The pricing page header still says "as of October 19, 2026"; Inkling 2 Small rows are marked New. Use the live table numbers above, not an older snapshot.
Tinker pricing page (official)
Are the weights open yet?
No. The model card lists Tinker Terms of Use. The November 5 blog said full weights would ship under Apache 2.0 "next week" after external safety testing and red-teaming.
As of November 9, 2026 Taipei, that release has not landed on a public weight card we could verify. Do not plan local GPU installs from this page. When weights open, revisit hosting with an official recipe; until then, use Tinker.
Inkling 2 Small model card (official)
What context lengths can you buy on Tinker?
Keep two numbers separate:
- Model card ceiling: up to 1M tokens
- Tinker SKUs today: 64K and 256K (training); serverless Beta lists 256K
A 1M context on the card does not mean a 1M purchasable Tinker tier in the table we checked.
What this page does not cover
We are not inventing VRAM GB tables, local GPU recipes, or China-cloud reseller markups. Weights are still closed.
We also are not dumping the full vendor benchmark suite. The AA Index 31 cite is the blog's published figure, labeled vendor-reported.
Related reading
- Images 2.6 / GPT-Image-2.6 (another Nov 2026 AI-tool short page)
- Kimi K3.1 Pro weights hosting (open weights case; contrast with Inkling's closed state)
- Particle-1 token efficiency (AI tool pricing checklist)
- SWE-2 Flash (coding-model short page)
Sources (checked November 9, 2026, Taipei)
- https://thinkingmachines.ai/news/inkling-2-small/
- https://thinkingmachines.ai/model-card/inkling-2-small/
- https://thinkingmachines.ai/tinker/pricing/
- https://tinker-docs.thinkingmachines.ai/tinker/models/models_and_pricing/
FAQ
What is Inkling 2 Small?
Inkling 2 Small is Thinking Machines Lab's November 5, 2026 second-generation MoE model: 594B total parameters, 17B active, text and image inputs, variable effort, and a model-card context window up to 1M tokens. It does not support audio. Distribution today is via Tinker for fine-tuning, sampling, and a playground.
How much does Inkling 2 Small cost on Tinker?
Per the Tinker pricing page (checked November 9, 2026 Taipei; table header dated October 19, 2026 with Inkling 2 Small rows marked New): 64K and 256K share prefill $0.30, cached prefill $0.03, and sample $1.00 per 1M tokens. Train is $1.50 per 1M on 64K and $3.00 on 256K. Serverless inference Beta (Inkling family only) lists 256K at input $0.30, cached input $0.03, output $1.00 per 1M. Storage is $0.10 per GB per month.
Are the Inkling 2 Small weights open yet?
No. The model card lists Tinker Terms of Use. The November 5 blog said full weights would follow under Apache 2.0 after external safety testing and red-teaming. As of November 9, 2026 Taipei, weights remain closed; Hugging Face returned no public card for the model ID we checked.
What context lengths can you buy on Tinker?
The model card states a context window up to 1M tokens. Tinker currently sells 64K (thinkingmachines/Inkling-2-Small) and 256K training (thinkingmachines/Inkling-2-Small:peft:262144). Serverless Beta lists the 256K SKU. Do not treat the 1M model-card ceiling as a purchasable Tinker tier today.
What does the AA Intelligence Index score of 31 mean here?
The official blog reports an Artificial Analysis Intelligence Index score of 31 for Inkling 2 Small. That figure is vendor-reported, not our benchmark. This page does not reproduce the full eval tables.
Key takeaways
- Inkling 2 Small is Thinking Machines Lab's second-generation model, released November 5, 2026.
- It is a 48-layer decoder-only MoE with 594B total parameters and 17B active, text plus image input, and a model-card context window up to 1M tokens.
- It ships on Tinker with closed weights under Tinker Terms of Use.
SourcesInkling 2 Small announcement (official) · Tinker pricing · Models & Pricing · Inkling 2 Small model card (official)
Last updated
Cite this pageWhat Is Inkling 2 Small? Tinker Pricing and Closed Weights (2026) — https://www.agentcomm.net/articles/inkling-2-small/