On November 5, 2026, Thinking Machines Lab released Inkling 2 Small, a 594-billion-parameter Mixture-of-Experts model with 17 billion active parameters. It runs on Tinker for sampling and fine-tuning with vision and up to 1M context, scores about 31 on the Artificial Analysis Intelligence Index, and is planned for Apache 2.0 weights after safety testing.
Last updated: November 6, 2026
Table of contents
- What shipped on November 5, 2026
- Specs, family comparison and pricing
- How to try and fine-tune on Tinker
- Open weights and safety path
- Limits and practical notes
- Frequently asked questions
- Sources
What shipped on November 5, 2026
Thinking Machines Lab’s Introducing Inkling 2 Small post launches the first model in its second generation. The company describes a Mixture-of-Experts system with 594 billion total parameters and 17 billion active, image inputs, and variable effort. On the Artificial Analysis Intelligence Index it reports a score of 31 (the model-card comparison table lists AA Index v5.0 at 30.6), with gains called out in agentic coding, vision, knowledge work, and reasoning versus prior Inkling models.
Access starts on Tinker: closed weights for sampling and fine-tuning now, with downloadable checkpoints planned when weights open. The announcement and model card frame Inkling 2 Small as a general-purpose multimodal model for developers building agents, coding assistants, chatbots, and retrieval systems—not as a chat-first consumer app. For related open MoE and coding coverage on FreeAnswers, see Kimi K3.1 Pro, SWE-2 Flash (built on Inkling-Small), and Muse Spark Open.
Specs, family comparison and pricing
Figures below come from the model card, the Inkling family page, and Tinker’s Models & Pricing / pricing pages (checked this run).
| Item | Inkling 2 Small | Inkling-Small | Inkling |
|---|---|---|---|
| Total / active params | 594B / 17B | 276B / 12B | 975B / 41B |
| Architecture type (Tinker) | Hybrid + Vision | Hybrid + Audio + Vision | Hybrid + Audio + Vision |
| Tinker ID (64K) | thinkingmachines/Inkling-2-Small |
thinkingmachines/Inkling-Small |
thinkingmachines/Inkling |
| Prefill / cached / sample / train (64K) | $0.30 / $0.03 / $1.00 / $1.50 | $0.58 / $0.116 / $1.44 / $1.73 | $1.87 / $0.374 / $4.68 / $5.61 |
| Train at 256K (same prefill/sample) | $3.00 | $3.47 | $11.23 |
| Model-card context | Up to 1M tokens | — | — |
| Weights at Inkling 2 Small launch | Closed on Tinker; Apache 2.0 planned after safety testing | Open (prior release) | Open (prior release) |
Architecture detail from the model card: a 48-layer decoder-only MoE that routes each token to 8 of 384 experts, hybrid local/global attention, shared image/text hidden space, and MXFP8 / NVFP4 numerics. The announcement adds a 292M-parameter vision encoder based on Perception Encoder, with 20×20 patches and 2×2 spatial pooling. Image inputs should stay between about 40px and 4096px per dimension for best results. For other recent MoE checkouts, see ERNIE 5.5 Flash Base, LFM3, and Nemotron 3.5.
How to try and fine-tune on Tinker
Tinker is Thinking Machines’ hosted training and sampling API. You keep the data, loss, and loop; Tinker runs distributed LoRA training and sampling. MoE models are priced by active parameters.
- Create a Tinker account and open Tinker Playground to try Inkling 2 Small interactively (linked from the announcement).
- Install the Tinker client / Cookbook stack from the Tinker Cookbook and follow the Get Started flow in the Tinker docs.
- For sampling, create a sampling client with base model
thinkingmachines/Inkling-2-Small. For 256K PEFT training usethinkingmachines/Inkling-2-Small:peft:262144. Serverless inference (beta) usesthinkingmachines/Inkling-2-Small:peft:262144:sampling-nvfp4. - For fine-tuning, create a LoRA training client on the same base model string, then use
forward_backward,optim_step, andsampleas in the docs. Installtml-rendererswhen you need the model’s chat/vision renderers (model card). - Watch spend: cached prefill is discounted; forward-only training calls bill at the train rate; checkpoints cost $0.10/GB/month. When weights open, Thinking Machines says trained checkpoints become downloadable with the weight release.
Variable effort is trained in: the announcement describes sampling effort levels with a length penalty so you can trade thoroughness for latency and cost. Prefer lower effort for interactive loops and higher effort for hard agentic or vision-in-the-loop jobs.
Open weights and safety path
Inkling 2 Small ships closed-weight on Tinker first. The announcement states an intent to publish full weights under Apache 2.0 after external safety testing and red-teaming—“next week” in the Nov 5 wording—consistent with the lab’s safe path to open weights essay: staged access (API, hosted fine-tuning, vetted research) before irreversible public weights when evidence and ecosystem readiness support it.
The model card and announcement summarize safety work across everyday interaction, CBRN, cyber, and alignment under long-horizon agentic RL. Early RL checkpoints showed misbehavior under hard tasks; the lab describes midtraining filters, an alignment reward, targeted environments, and on-policy self-distillation, plus synchronous and asynchronous tool-call monitors. Treat the planned open release as contingent on those external tests completing—not as a guaranteed calendar date.
Limits and practical notes
Official limitations: no audio (unlike Inkling and Inkling-Small), weaker chat-style presentation than some peers, and occasional over-refusal on mental-health-adjacent prompts. Knowledge is bounded by the training cutoff. The model card recommends human oversight for high-stakes medical, legal, or safety-critical use, plus application-layer filtering and monitoring for open deployments.
Benchmark tables on the announcement and model card are broad; cite them from those pages rather than third-party roundups. Artificial Analysis is named by Thinking Machines for the Index score near 31. For adjacent FreeAnswers guides on multimodal APIs, see Ling-3.2-flash and MiniMax M3.1 Flash.
Frequently asked questions
What is Inkling 2 Small?
Inkling 2 Small is Thinking Machines Lab’s second-generation model, announced November 5, 2026. It is a multimodal Mixture-of-Experts transformer with 594 billion total parameters and 17 billion active per token, accepting text and image inputs and producing text. It is available on Tinker for inference and fine-tuning while weights remain closed pending further safety testing.
How do I try Inkling 2 Small?
Thinking Machines says you can try it today on Tinker Playground. Developers can also sample and fine-tune through the Tinker API using the model ID thinkingmachines/Inkling-2-Small (and the 256K PEFT / serverless variants listed on the Models & Pricing docs). The Tinker Cookbook and Get Started docs cover the training and sampling clients.
How much does Inkling 2 Small cost on Tinker?
Per the Tinker Models & Pricing docs (USD per 1M tokens): at 64K context, prefill is $0.30 ($0.03 cached), sample $1.00, and train $1.50. At 256K context, prefill and sample stay the same while train rises to $3.00. Serverless inference (beta) for Inkling 2 Small lists $0.30 input, $0.03 cached input, and $1.00 output at 256K. Checkpoint storage is $0.10 per GB per month.
When will Inkling 2 Small open weights be released?
At launch the weights are closed on Tinker. The November 5 announcement says Thinking Machines intends to release full weights under an Apache 2.0 license next week after external safety testing and red-teaming, following the lab’s safe path to open weights approach. The family page similarly says open weights are intended in the coming weeks, pending additional safety testing.
How does Inkling 2 Small differ from Inkling and Inkling-Small?
Per the Inkling family page and Tinker pricing tables: Inkling is 975B total / 41B active with Hybrid + Audio + Vision; Inkling-Small is 276B / 12B with the same modality mix; Inkling 2 Small is 594B / 17B with Hybrid + Vision (no audio). Inkling 2 Small is positioned as the most capable Inkling model so far, with stronger agentic coding and vision results than the first generation.
What is the context window and architecture?
The model card lists a context window of up to 1 million tokens, MXFP8 and NVFP4 numerics, and a 48-layer decoder-only MoE that routes each token to 8 of 384 experts, with hybrid local/global attention and a vision encoder. Tinker currently prices 64K and 256K training/sampling tiers for this model; use the matching Tinker ID for the context length you need.
Sources
- Thinking Machines Lab — Introducing Inkling 2 Small (Nov 5, 2026)
- Thinking Machines Lab — Inkling 2 Small model card
- Thinking Machines Lab — Inkling family page
- Tinker docs — Models & Pricing
- Thinking Machines Lab — Tinker pricing
- Thinking Machines Lab — A Safe Path to Open Weights