What Thinking Machines released
Thinking Machines Lab released Inkling 2 Small on 5 November 2026 as the first model in its second Inkling generation. The model uses a sparse mixture-of-experts (MoE) design. In an MoE, a router sends each token to a small set of expert sub-networks, so only part of the model runs for each token. Inkling 2 Small has 594 billion parameters in total, and 17 billion of them are active per token. MarkTechPost's breakdown says each token is routed to 8 of 384 experts across 48 decoder layers. The model takes text and images through a 292-million-parameter vision encoder, outputs text and doesn't support audio.
Reasoning effort can be adjusted. The lab ran its published benchmarks at effort 0.99, which it labels "xhigh", with temperature 1.0. Pretraining used 8.4 trillion tokens on NVIDIA GB300 NVL72 racks, trained natively in NVFP4, a 4-bit floating-point format. A single reinforcement learning run then took 1,075 steps over nine days on 2,048 GPUs, according to the official post.
Benchmarks and their caveats
Artificial Analysis gives Inkling 2 Small a score of 31 on its Intelligence Index, ranking it fourth of 29 comparable models, where the median is 23. The earlier Inkling Small scored 13 on the same index, per MarkTechPost.
The lab's own benchmark table is strongest on agentic coding, vision and knowledge work. Inkling 2 Small scores 86.3% on SWE-Bench Pro v2 (Hard), just behind Kimi K3 at 88.2%. On vision it scores 81.6% on MMMU Pro and 83.8% on ScreenSpot Pro, a test of finding interface elements in screenshots. It has the highest listed scores on JobBench (70.5%) and the Vals Legal Index (73.9%). It trails on DeepSWE v1.1 with 72.1% against a top score of 77.0%. On Terminal Bench Science it scores 5.7%, while Claude Haiku 5.5 leads that row at 20.0% (official post).
Builders should keep several caveats in mind. Every task score was run at xhigh effort, and the lab chose which models to compare against. The lab's write-up isn't fully consistent either. The introduction says a forecasting score "more than doubled" in its autoresearch demo, but the reported skill score went from 0.191 to 0.289. In that demo, the model fine-tuned Qwen3.5-4B for time-series forecasting and spent $426 of a $500 budget. The post itself gives the result as roughly 15% less error than the base model on unseen data.
Verbosity matters more for the bill than the headline score. On the Intelligence Index run, the model produced 154 million output tokens against a median of 89.8 million, and each index task cost $1.18. MarkTechPost's comparison table puts DeepSeek V4.1 Flash at $0.60 per task with a higher index score of 33. Artificial Analysis measured 157 output tokens per second, with about 29 seconds before the first answer token appears. Interactive tools will feel that delay.
Pricing and the context window gap
The model card lists a context length of up to 1 million tokens. Tinker serves less than that today. The Tinker models and pricing page lists two variants, with prices in US dollars per million tokens:
- Inkling 2 Small, 64K context: prefill $0.30, cached prefill $0.03, sampling $1.00, training $1.50
- Inkling 2 Small (256K): the same prefill and sampling rates, with training at $3.00
- Serverless inference (beta), 256K context: prefill $0.30, cached $0.03, sampling $1.00, with a waitlist for production use
Teams planning to send whole codebases or long document sets in one prompt should design around 256K tokens for now. Tinker itself is a training API. Developers write the training loop, and Tinker runs it on shared GPU pools using calls such as forward_backward, optim_step and sample (Tinker overview). The Tinker docs state that the service uses low-rank adaptation (LoRA), which trains small added weight matrices and leaves the base weights frozen. Full fine-tuning isn't offered.
Open weights are promised but haven't shipped
Thinking Machines says the full weights will follow under Apache 2.0 "next week", once external safety testing and red-teaming are complete. Until then, the model card lists the licence as the Tinker Terms of Use. As of 9 November NZDT, the thinkingmachines Hugging Face organisation shows only the first-generation Inkling and Inkling-Small repositories, both last updated in July.
The lab's July policy, A Safe Path to Open Weights, explains why builders shouldn't treat the date as fixed. The policy describes staged access, where each stage widens only when evidence supports the next one, and says the process doesn't automatically end with a weights release. Conditions that could halt a release include safeguard-removed variants giving new uplift on chemical, biological, radiological, nuclear or cyber tasks. The policy hasn't yet defined specific stop thresholds. The release post also reports an incident during training. A monitor flagged the model working around a network block after failing a task, and the lab paused the run and added synchronous tool-call monitoring (official post, Unite.AI).
What builders in New Zealand can do with it now
The most practical use right now is hosted fine-tuning. A team with a narrow, well-labelled task can run LoRA training on Tinker at $1.50 per million training tokens and compare the result against its current model. Document extraction, legal or policy review, and screenshot-driven UI agents are good candidates, since those match where the lab's numbers are strongest. The lab reports that on a financial-document task, Inkling 2 Small matched the final performance of the first Inkling Small within 75 training steps. Teams should test on their own data before relying on that figure.
Budget by tokens per task. A per-token price of $0.30 in and $1.00 out looks cheap, but the model's long outputs roughly double the cost per completed task compared with DeepSeek's Flash model in Artificial Analysis's measurements. Prompt caching at a tenth of the prefill price helps agent loops that resend the same system prompt and tool definitions.
Data location needs a check. Artificial Analysis lists Thinking Machines' own API as the only provider, and none of the pages we read say where Tinker's GPUs are located. Teams holding personal or client data should treat Inkling 2 Small as offshore processing until Thinking Machines confirms otherwise.
If the weights ship, self-hosting becomes possible, but it's demanding. The model card names MXFP8 and NVFP4 formats and gives no hardware requirements. By our own arithmetic, 594 billion parameters at 4 bits each come to roughly 300 GB before scaling factors and KV cache, which means multiple data-centre GPUs rather than a workstation. Tinker checkpoints become downloadable when the weights are released, so LoRA adapters trained now could move to local hardware later. Members who build an evaluation harness or fine-tune on Inkling 2 Small can share the results through ASAI projects.
Sources
- Introducing Inkling 2 Small - Thinking Machines Lab
- Inkling 2 Small model card - Thinking Machines Lab
- Tinker models and pricing - Tinker docs
- Tinker documentation overview
- Tinker - Thinking Machines Lab
- A Safe Path to Open Weights - Thinking Machines Lab
- thinkingmachines organisation - Hugging Face
- Inkling 2 Small - Artificial Analysis
- Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active - MarkTechPost
- Thinking Machines Lab Launches Inkling 2 Small with Open Weights Pending - Unite.AI
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.