Inkling 2 Small: Launch and confirmed availability
Image generated by the Inferama Editorial Team
01

What was launched and when

Thinking Machines Lab has released Inkling 2 Small, a general-purpose multimodal model that accepts text and images as input and generates text as output. The official launch date is November 5, 2026, according to the model's technical card.

The model is designed for developers building AI-powered applications, including agentic systems, coding assistants, chatbots, and retrieval-augmented generation (RAG) systems. It supports English and other languages, as well as multiple programming languages.

02

Architecture and technical specifications

Inkling 2 Small is a 48-layer decoder-only autoregressive transformer with a sparse Mixture-of-Experts (MoE) backbone. Each token is routed to 8 of 384 experts, enabling significant computational efficiency.

The model has 594 billion total parameters, of which 17 billion are active per token. Attention is hybrid, combining local and global layers. It is natively multimodal: images are encoded via a vision transformer and projected into a shared hidden space with text for joint processing by the decoder.

It supports MXFP8 and NVFP4 numerics. The context window declared in the technical card is up to 1M tokens, although actual availability on the Tinker platform varies by use case.

Inkling 2 Small technical specifications

FeatureDetail
Total parameters594B
Active parameters per token17B
ArchitectureSparse MoE (8 of 384 experts)
Layers48
Input modalitiesText and image
Output modalityText
Context window (card)Up to 1M tokens
Numeric supportMXFP8, NVFP4
03

Variants, modalities, and access

The model is available through the Tinker platform, which provides API access for fine-tuning and sampling. Developers can start by installing the `tml-renderers` package and using the Tinker Cookbook.

Tinker offers two main configurations for Inkling 2 Small: one with a 64K token context and another with 256K tokens (identified by the `:peft:262144` suffix). The 256K configuration has a higher training cost, but prefill and sampling prices are the same as the 64K version.

Additionally, there is a beta serverless inference option with a specific identifier (`:sampling-nvfp4`), currently available only for Inkling, Inkling-Small, and Inkling 2 Small. For production use, interested parties must request access through a waitlist.

04

Pricing and operational costs

Prices on Tinker are calculated per million tokens. For Inkling 2 Small, the prefill (input) cost is $0.30, with a discount to $0.03 for cached tokens. The sampling (output) cost is $1.00 per million tokens.

Training costs $1.50 per million tokens in the 64K configuration and $3.00 in the 256K configuration. Checkpoint storage is billed at $0.10 per GB per month.

In the beta serverless inference modality, prices are $0.30 for input, $0.03 for cached input, and $1.00 for output, with a maximum context of 256K tokens.

Inkling 2 Small pricing on Tinker (per million tokens)

ModalityContextPrefill (USD)Cached Prefill (USD)Sampling (USD)Training (USD)
Standard64K0.300.031.001.50
Extended256K0.300.031.003.00
Serverless (beta)256K0.300.031.00N/A
05

Evaluations and performance

The technical card publishes a series of results on third-party benchmarks. On the Artificial Analysis Intelligence Index v5.0, Inkling 2 Small scores 30.6. On DeepSWE v1.1, it reaches 72.1%, and on SWE-bench Pro v2 Hard, 86.3%.

These results are published by the provider and have not been independently reproduced by Inferama. Performance may vary depending on configuration, language, and specific application domain. Developers are recommended to conduct their own evaluations before deploying the model in high-risk or safety-critical contexts.

The model was evaluated for safety prior to release, including everyday interaction and dangerous capability testing. Mitigations were applied to reduce risks, and the model was found to be materially below frontier capabilities in terms of loss of control.

06

Limitations and safety considerations

Inkling 2 Small may exhibit limitations common to foundation models, including hallucinations (generating plausible but factually incorrect content), occasional failures to follow instructions precisely, and degraded performance in long multi-turn conversations.

Trained on web-derived and synthetic data, the model may reflect demographic, cultural, or linguistic biases present in the training data. Its knowledge is limited to information available up to its training cutoff and does not reflect subsequent events.

Developers are recommended to apply human oversight and review for outputs used in high-risk or safety-critical contexts. The model should not be deployed in domains such as medicine, law, or safety-critical decision-making without additional fine-tuning, domain-specific validation, and human oversight.

07

Keep exploring

Sources consulted

03

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.

Submit a correction