Inkling 2 Small

Inkling 2 Small is a general-purpose multimodal model from Thinking Machines Lab. It takes text and image inputs and generates text. It is built for agentic and tool-use systems, coding assistants, chatbots and retrieval-augmented generation, as well as general conversation, instruction following and visual tasks.

Languages: English, with general multilingual capabilities across other languages.

Model properties

Model type	Multimodal autoregressive transformer
Architecture	48-layer decoder-only transformer with a sparse Mixture-of-Experts backbone (8 of 384 experts per token); hybrid local and global attention; images encoded by a vision transformer and processed jointly with text
Parameters	594B total, 17B active
Precision	NVFP4 (the model's native training precision)
Input	Text (UTF-8) and images (each dimension ideally 40–4096 px)
Output	Text (UTF-8)
Context window on DeepInfra	262,144 tokens

Features on DeepInfra

• Reasoning with adjustable effort: set reasoning_effort to none, low, medium, high or xhigh. Reasoning is on by default; none turns it off.
• Tool calling, including tool_choice set to auto, required, none or a named function.
• Structured output with json_object and json_schema.
• Vision: pass images as image_url content parts.
• Prompt caching: repeated prefixes are served from cache and billed at the cached-input rate.

Training

The training data covers a broad variety of content types, including text, images and video. It was drawn from publicly available sources, acquired from third parties, generated internally, or synthetically generated or augmented. Curation included deduplication and filtering of low-quality data.

Evaluations

Selected results reported by Thinking Machines Lab.

Category	Benchmark	Inkling 2 Small
Index	Artificial Analysis Index (v5.0)	30.6
Index	Vals Index (v2.2)	56.1%
Coding	SWE-Bench Pro v2 (Hard)	86.3%
Coding	DeepSWE v1.1	72.1%
Coding	Vibe Code Bench v1.1	81.6%
Search and tools	BrowseComp (with compaction)	90.2%
Search and tools	DeepSearchQA (with compaction)	90.9%
Search and tools	MCP Atlas (full / public)	85.4% / 85.0%
Knowledge work	GDPVal-AA v2.1	1630
Finance	Finance Agent v2.1	67.5%
Visual perception	ScreenSpot Pro (original / with Python)	83.8% / 92.0%
Visual perception	OmniDocBench v1.6	94.6%
Visual reasoning	MMMU Pro (Standard 10)	81.6%
STEM reasoning	HLE (text only)	44.1%
STEM reasoning	ARC-AGI-2	81.7%
Long context	AA-LCR v1.1	83.3%
Long context	LongBench v2	70.6%
Instruction following	ComplexConstraints v1.0	44.5%
Multilinguality	Global-MMLU-Lite	91.0%
Factuality	AA-Omniscience v1.1 (non-hallucination rate)	73.3%

Safety

Thinking Machines Lab ran safety evaluations before release, covering everyday human-AI interaction (sycophancy, manipulation, psychological harm, over-refusal) and dangerous capabilities (CBRN, cyber, loss of control), across both text and image inputs. They concluded the model does not present risk of material uplift beyond what is already available in the open-weight ecosystem.

Limitations

Like other foundation models, Inkling 2 Small can hallucinate, occasionally fail to follow instructions precisely, and degrade in long multi-turn conversations. It may reflect biases in its training data and perform unevenly across languages and domains. Its knowledge stops at its training cutoff.

For high-stakes uses, evaluate the model on your own task, add application-level safeguards, and keep human review in the loop. Avoid medical, legal or safety-critical decision-making without domain-specific validation and human oversight.

License

Apache 2.0. Use is also subject to Thinking Machines Lab's Acceptable Use Policy. See also the training data documentation.