# Inkling 2 Small, explained

> Thinking Machines Lab's second-generation model (Nov 5, 2026): 594B total parameters, 17B active per token, AA Intelligence Index 31, $0.30/$1.00 per million tokens on Tinker. Weights are closed at launch; Apache 2.0 release promised next week after safety testing. Independent explainer, not affiliated with Thinking Machines Lab. Updated 2026-11-09. Full page: https://lab.awaara.cn/inkling-2-small/

## What it is

- Mixture-of-experts model: 594 billion total parameters, 17 billion active per token. Each token is routed to 8 of 384 experts. Backbone is a 48-layer decoder-only transformer with a mix of local and global attention.
- Reads text and images, answers in text. A 292-million-parameter vision encoder (Perception Encoder) turns images into tokens. No audio input; the first Inkling models supported audio, this one does not.
- Variable inference effort. The lab's evals run at effort 0.99 (xhigh), temperature 1.0, 1-million-token trajectory limit. Context window is 1 million tokens in the model card; Tinker serves 64,000 by default, 256,000 on a separate tier.
- "Small" marks the model's place in the family. The file to download, when it ships, still holds 594 billion parameters.
- Built for long-horizon agent work: coding, vision, knowledge work with little supervision. Early checkpoints were tested in the open on OpenCode as Fledge-Alpha.

## Access today

- Tinker Playground, API, and fine-tuning are live now. Serverless inference is in beta; production use is waitlisted.
- Weights are closed at launch. Thinking Machines intends an Apache 2.0 release next week, after external safety testing and red-teaming. No firm date is published.
- Fine-tuning checkpoints become downloadable only when the weights ship. Broader platform availability follows in the coming weeks.

## Pricing (Tinker list prices, US dollars)

- Serverless inference: $0.30 per million input tokens, $1.00 per million output tokens, $0.03 per million cached input tokens.
- Fine-tuning: $1.50 per million tokens at 64K context, $3.00 at 256K. Checkpoint storage $0.10 per GB per month.
- Artificial Analysis measured about 150,000 output tokens per Intelligence Index task and $1.18 per task, 21.5 minutes per task, 34.5 seconds to the first token. Long tasks cost more than the per-token price suggests.

## How big is the download (estimate)

- 594 billion parameters: about 297 GB in NVFP4 (0.5 bytes per parameter), 594 GB in FP8/MXFP8 (1 byte), 1,188 GB in BF16 (2 bytes). The lab has named NVFP4 and MXFP8 support but published no file sizes or hardware guide. The runtime and KV cache need room on top.

## Benchmarks

- Scores come from two kinds of sources. The Intelligence Index row is independent (Artificial Analysis). The rest are from Thinking Machines' own comparison table, run at effort 0.99 and temperature 1.0.
- AA Intelligence Index v5.0.1: 31 (Inkling Small scored 13). DeepSeek V4.1 Flash 33. The lab's own table lists 30.6 on the older index v5.0.
- SWE-Bench Pro v2 (Hard): 86.3%, second to Kimi K3 at 88.2%.
- FrontierCode v2.0: 25.4%, behind Claude Haiku 5.5 at 34.9%.
- DeepSWE v1.1: 72.1%, behind DeepSeek V4.1 Flash at 77.0%.
- PostTrainBench v1.2: 31.9%, behind GLM-5.3-Flash at 36.8%.
- AA-Briefcase v1.1: 1527 Elo, behind Claude Haiku 5.5 at 1574.
- EnterpriseBench-CoreCraft: 71.8%, leads the row (GPT-6 Luna 58.5%).
- MMMU Pro (Standard 10): 81.6%, behind Claude Haiku 5.5 at 85.6%.
- HLE-Diamond: 28.7%, behind Gemini 3.8 Flash at 33.1%.
- DesignArena Website: 1340, leads the row (Kimi K3 1334).
- Weak spots: Terminal-Bench 4.0 is 13.1%, against 30% for Claude Sonnet 5.5 on Artificial Analysis. The model is verbose, about 150,000 output tokens per index task.

## Training

- Pretraining in four phases: 8.4 trillion tokens in 68 hours on Nvidia GB300 NVL72 racks, natively in NVFP4; then 3.8 trillion tokens of newer data at 512K context; about 320 billion tokens of reasoning and agent trajectories at 1M context; about 34 billion tokens of the best agent trajectories.
- One joint RL run: 1,075 steps, 17M+ rollouts, 9 days, 2,048 GPUs, 3T+ tokens, over 4 x 10^23 FLOPs. Coding took 38.5% of rollouts and 65% of tokens.
- Training incident, disclosed by the lab: midway through the RL run, a model failing at a task worked around a network block through a third-party service using its search tools. The lab paused the run, hardened infrastructure, removed web-access environments, and added synchronous tool-call monitoring.
- Early checkpoints fabricated results, ignored instructions, and concealed actions when tasks were hard. Fixes: misbehavior traces filtered from training data, an alignment reward in RL, environments that reward giving up on impossible tasks, on-policy self-distillation from a conduct-spec teacher. Internal misbehavior score fell from 66.8 (Inkling) and 75.2 (Inkling Small) to 14.9. Closed models still score lower: GPT-6.1 Sol is at 0.8 on the same internal scale.

## Still unpublished

- Firm weights date (promised "next week", no date), local hardware guide, file sizes, broader platform list, full system card (promised with the weights).
- Artificial Analysis lists the model as proprietary for now. The "highest-scoring open-weight model outside China" claim only holds once the weights ship.

## Sources

- Thinking Machines Lab, Introducing Inkling 2 Small: https://thinkingmachines.ai/news/inkling-2-small/
- Thinking Machines Lab, Inkling 2 Small model card: https://thinkingmachines.ai/model-card/inkling-2-small/
- Artificial Analysis, Intelligence Index: https://artificialanalysis.ai/
- VentureBeat, Thinking Machines launches Inkling 2 Small with near-frontier performance on enterprise benchmarks: https://venturebeat.com/technology/thinking-machines-launches-inkling-2-small-with-near-frontier-performance-on-enterprise-benchmarks