Inkling-2-Small
A 594B-parameter sparse Mixture-of-Experts model with 17B active per token, built for agentic work, coding, vision and retrieval-augmented generation. Reads text, images and PDFs; returns text. Reasoning runs by default with adjustable effort and can be switched off. Supports tool calling, JSON and schema-constrained output, and prompt caching across a window of over one million tokens.

- Tools
- Vision
- PDF input
- Reasoning
- Reasoning
- On by default
reasoning_effort- none
- low
- medium
- high
Can be turned off
- Tool choice
requiredfunction
Use Inkling-2-Small in VS Code, Cursor and Antigravity
DEVUP Code, the DEVUP AI extension, is free to install. Requests are paid from your DEVUP balance.
Inkling 2 Small
Inkling 2 Small is a general-purpose multimodal model from Thinking Machines Lab. It reads text, images and PDF documents, and writes text. The design targets agentic and tool-driven systems, coding assistants, chat products, retrieval-augmented generation, instruction following and visual understanding.
The model is a sparse Mixture-of-Experts transformer: 594 billion parameters in total, 17 billion active for any single token. Each token is routed to 8 of 384 experts across a 48-layer decoder-only stack that mixes local and global attention. Images pass through a 292-million-parameter vision encoder and are processed jointly with the text stream.
It works in English and carries general multilingual ability across other languages, including multiple programming languages.
At a glance
| Model type | Multimodal autoregressive transformer |
| Total parameters | 594B |
| Active parameters per token | 17B |
| Layers | 48, decoder-only |
| Expert routing | 8 of 384 per token |
| Attention | Hybrid local and global |
| Input | Text, images, PDF |
| Output | Text |
| Maximum context size | 1,048,576 tokens |
| License | Apache 2.0, plus the publisher's acceptable use policy |
Maximum context size counts input and output together. A request that reserves the whole window for output leaves no room for a prompt, so size the two against each other.
Getting started
curl https://api.devupai.com/v1/chat/completions \
-H "Authorization: Bearer $DEVUP_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "thinkingmachines/Inkling-2-Small",
"messages": [
{ "role": "user", "content": "Summarise the trade-offs of sparse Mixture-of-Experts routing." }
]
}'Reasoning
Reasoning runs by default, and the deliberation arrives as readable text in the response's reasoning field alongside a count of the tokens it consumed.
Five effort levels are documented:
| Level | Behaviour |
|---|---|
none | Reasoning switched off |
low | Short deliberation |
medium | Balanced |
high | Extended deliberation |
xhigh | Maximum deliberation |
curl https://api.devupai.com/v1/chat/completions \
-H "Authorization: Bearer $DEVUP_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "thinkingmachines/Inkling-2-Small",
"reasoning_effort": "xhigh",
"messages": [
{ "role": "user", "content": "A train leaves at 14:05 travelling 82 km/h. A second leaves the same station at 14:40 travelling 118 km/h on the same track. When does the second catch the first?" }
]
}'Higher levels spend more tokens before answering and take longer to return. Lower levels suit classification, extraction and routing; higher levels suit multi-step mathematics, planning and difficult debugging.
Two ways to switch deliberation off entirely, both producing a response with no reasoning:
{ "reasoning_effort": "none" }{ "reasoning": { "enabled": false } }Switch it off whenever the answer is short and mechanical. With reasoning on, deliberation consumes the output budget before the answer begins, so a tight max_tokens can cut a reply off mid-thought. Either raise the budget or lower the effort.
Vision
Images are passed as image_url content parts. Each side of an image works best between 40 and 4096 pixels.
curl https://api.devupai.com/v1/chat/completions \
-H "Authorization: Bearer $DEVUP_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "thinkingmachines/Inkling-2-Small",
"messages": [{
"role": "user",
"content": [
{ "type": "text", "text": "Which row of this invoice has an arithmetic error?" },
{ "type": "image_url", "image_url": { "url": "https://example.com/invoice.png" } }
]
}]
}'A data URI works in place of a hosted link:
{ "type": "image_url", "image_url": { "url": "data:image/png;base64,iVBORw0KGgo..." } }Several images can appear in one message, each as its own content part. PDF documents are accepted the same way. Audio and video are not.
Tool calling
Functions are declared in tools, and tool_choice accepts auto, required, none or a named function.
curl https://api.devupai.com/v1/chat/completions \
-H "Authorization: Bearer $DEVUP_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "thinkingmachines/Inkling-2-Small",
"messages": [{ "role": "user", "content": "How many open tickets does the billing team have?" }],
"tools": [{
"type": "function",
"function": {
"name": "count_tickets",
"description": "Count open tickets for a team",
"parameters": {
"type": "object",
"properties": {
"team": { "type": "string", "description": "Team name" },
"status": { "type": "string", "enum": ["open", "closed", "all"] }
},
"required": ["team"]
}
}
}],
"tool_choice": "auto"
}'Forcing one named function:
{ "tool_choice": { "type": "function", "function": { "name": "count_tickets" } } }Feed each result back as a tool message carrying the matching tool_call_id, then call again so the model can finish its answer.
Structured output
Both json_object and json_schema are accepted.
curl https://api.devupai.com/v1/chat/completions \
-H "Authorization: Bearer $DEVUP_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "thinkingmachines/Inkling-2-Small",
"messages": [
{ "role": "user", "content": "Extract the supplier, total and due date from: Nordwind GmbH, 4,820.50 EUR, payable by 2026-12-03." }
],
"max_tokens": 4096,
"response_format": {
"type": "json_schema",
"json_schema": {
"name": "invoice",
"strict": true,
"schema": {
"type": "object",
"properties": {
"supplier": { "type": "string" },
"total": { "type": "number" },
"currency": { "type": "string" },
"due_date": { "type": "string" }
},
"required": ["supplier", "total", "currency", "due_date"],
"additionalProperties": false
}
}
}
}'Keys may arrive in a different order from the schema; the values and types follow it. Give max_tokens enough room for the deliberation that precedes the object, or set reasoning_effort to none for pure extraction work.
Prompt caching
A repeated opening section of a conversation is served from cache and billed at the cached-input rate. Put the stable material first — system instructions, reference documents, few-shot examples — and the part that changes last. The longer the shared prefix, the larger the saving.
Long context
The window holds over a million tokens across input and output combined. That covers an entire codebase, a full contract set, or a long conversation with every document it refers to. Leave headroom for the answer, and for deliberation tokens when reasoning is on.
Choosing settings
| Task | Suggested configuration |
|---|---|
| Classification, routing, extraction | reasoning_effort: "none" |
| Chat and drafting | reasoning_effort: "medium" |
| Mathematics, planning, hard debugging | reasoning_effort: "high" or "xhigh" |
| Agent loops | tools plus tool_choice: "auto" |
| Machine-readable results | response_format with json_schema and a generous max_tokens |
| Screenshots, charts, scanned documents | image_url parts, reasoning_effort: "medium" or higher |
| Repeated long preambles | Stable content first, variable content last |
Limits to keep in mind
Like any model of this kind, Inkling 2 Small can state something false with confidence, miss part of a precise instruction, and lose coherence over very long multi-turn exchanges. Quality varies across languages and domains, and its knowledge ends at its training cutoff.
For consequential work, measure it on your own task before relying on it, add checks in your application, and keep a person in the loop. Medical, legal and safety-critical decisions need domain validation and human oversight.
Licensing
Released under Apache 2.0, with use also governed by the publisher's acceptable use policy.