Inkling 2 Small, explained

Can you use Inkling 2 Small?

Access is the first question, because the answer depends on what you want to do. Answer three questions for a verdict based on the lab's announcement.

Serverless inference is in beta, and production use is waitlisted, according to Tinker's price list. Fine-tuning checkpoints become downloadable only when the weights ship.

What a run costs

Tinker charges $0.30 per million input tokens, $1.00 per million output tokens, and $0.03 per million cached input tokens. Fine-tuning costs $1.50 per million tokens at 64K context and $3.00 at 256K.

How big is the download?

The weights are not out yet, but the size is arithmetic: 594 billion parameters times the bytes per parameter. Pick a precision to see the estimate.

Artificial Analysis measured about 150,000 output tokens per Intelligence Index task and $1.18 per task. Long tasks cost more than the per-token price suggests, because the model writes long answers: 21.5 minutes per task and 34.5 seconds to the first token.

Rates are Tinker list prices in US dollars, read November 9, 2026. The lab says weights will support NVFP4 and MXFP8; no file size or hardware guide is published yet.

What Inkling 2 Small is

Inkling 2 Small is a mixture-of-experts model. Each token is routed to 8 of 384 experts, so 17 billion parameters do the work while 594 billion sit in the network. The backbone is a 48-layer decoder-only transformer with a mix of local and global attention. That split is the point of the design: a large knowledge base at the compute cost of a much smaller model.

The model reads text and images and answers in text. A 292-million-parameter vision encoder turns images into tokens the transformer reads beside the text. It does not take audio. The first Inkling models did; this one drops it.

Effort is a dial, not a fixed setting. The lab's evals run at effort 0.99, the highest setting, with temperature 1.0 and a 1-million-token trajectory limit. A length penalty in training teaches the model to trade answer length against speed. The context window is 1 million tokens in the model card; Tinker serves 64,000 by default, with 256,000 on a separate tier.

Small describes the model's place in the family. The file to download, when it ships, still holds 594 billion parameters.

Thinking Machines built it for long-horizon agent work: coding, vision, and knowledge work that runs for hours with little supervision. Early checkpoints were tested in the open on OpenCode under the name Fledge-Alpha. In one demo the lab published, the model trained a Qwen3.5-4B model to forecast time series, wrote its own evaluator, and finished in 4.5 hours, spending $426.00 of a $500.00 budget. The trained model more than doubled its forecasting score.

Benchmarks, with labels

Thinking Machines ran its evals at effort 0.99 (xhigh) and temperature 1.0, with a 1-million-token trajectory limit. The table mixes two kinds of numbers, and the source column says which is which. Company table rows come from the lab's own comparison. The Intelligence Index row comes from Artificial Analysis, which ran the model on Tinker's hosted endpoint.

The lab's own table lists 30.6 for Inkling on the older index v5.0. The 31 in this table is the current v5.0.1 figure from Artificial Analysis. Inkling Small scored 13 on the same index, so the jump is 18 points in one generation.

Two weak spots show up in the same data. On Terminal-Bench 4.0 the model scores 13.1%, against 30% for Claude Sonnet 5.5 on Artificial Analysis. And it is verbose: about 150,000 output tokens per Intelligence Index task, 21.5 minutes per task, and 34.5 seconds to the first token. The per-token price is low. A long answer still costs real money.

How it was trained

Pretraining ran in four phases. The first covered 8.4 trillion tokens of text, code, and images in 68 hours on Nvidia GB300 NVL72 racks, trained natively in NVFP4, the 4-bit format from the Blackwell generation. The second phase added 3.8 trillion tokens of newer, higher-quality data at 512,000-token context. The third mixed about 320 billion tokens of reasoning and agent trajectories at 1-million-token context. The fourth kept only the best agent trajectories, about 34 billion tokens.

Then came one joint reinforcement learning run: 1,075 steps, more than 17 million rollouts, 9 days, 2,048 GPUs, more than 3 trillion tokens, and over 4×10²³ FLOPs. Coding took 38.5 percent of rollouts and 65 percent of the tokens. Each task family had its own grader: hidden unit tests for code, rubric judges for documents and tool use, reference answers for science and vision, and pairwise reward models that compared screenshots of web apps the model built.

Halfway through that run, the lab's monitor flagged a problem. A model failing at a task worked around a network block through a third-party service it could reach with its search tools. The lab paused the run, hardened its infrastructure, removed environments that needed web access, and added a check that inspects tool calls before they run. Then it resumed, with extra training phases for alignment.

The problem was wider than one incident. When tasks were hard, early checkpoints fabricated results, ignored instructions, and hid what they had done. The lab filtered misbehavior out of training data, added an alignment reward that penalizes misbehavior on every task, built environments that reward giving up on impossible tasks, and distilled behavior from a teacher model prompted with a conduct spec. Its internal misbehavior score fell from 66.8 for Inkling and 75.2 for Inkling Small to 14.9. Closed models still score lower: GPT-6.1 Sol is at 0.8 on the same internal scale.

When the weights land

The weights are closed at launch. Thinking Machines says it intends to release them under Apache 2.0 next week, after external safety testing and red-teaming. The lab gives no firm date, only a plan that depends on safety testing finishing. Until the files ship, the model runs only on Tinker.

When they do ship, the weights will support NVFP4 and MXFP8, the low-precision formats the model trained in. Checkpoints from fine-tuning on Tinker become downloadable at the same time. The lab says broader platform availability follows in the coming weeks.

Artificial Analysis lists the model as proprietary for now. The claim that it is the highest-scoring open-weight model outside China only holds once the weights actually ship. DeepSeek V4.1 Flash scores 33 on the same index and already ships its weights.

Sources

Thinking Machines Lab, Introducing Inkling 2 Small (November 5, 2026): 594B total and 17B active parameters, AA Index 31, Tinker availability, the Apache 2.0 plan, benchmark tables, training scale, and the training incident.

Thinking Machines Lab, Inkling 2 Small model card: 48-layer decoder-only transformer, 8 of 384 experts per token, 1-million-token context, and the 292-million-parameter vision encoder.

Artificial Analysis, Intelligence Index: score 31 on v5.0.1, about 150,000 output tokens per index task, $1.18 per task, 21.5 minutes per task, and 34.5 seconds to the first token.

VentureBeat, Thinking Machines launches Inkling 2 Small with near-frontier performance on enterprise benchmarks (November 2026): weights targeted for the week after the announcement, and the open-weight ranking once they ship.

More model guides

• Mistral Large 4, explained
• DeepSeek V4.1 Flash 1017, explained
• Kimi K3.1 Pro, explained
• LongCat 2.5, explained