Inkling Small

Compact multimodal model with variable thinking effort and 1M context

Try now

read docs

About model

Inkling-Small is Thinking Machines Lab's compact open-weights model, trained with the same recipe as Inkling at roughly a quarter of its size: a Mixture-of-Experts transformer with 276B total parameters and 12B active per token. Like its larger sibling, it reasons natively over text, images, and audio in a shared hidden space, varies its thinking effort to balance cost against quality, and supports a context window of up to 1M tokens. The smaller footprint translates into faster inference and lower serving cost, and Thinking Machines Lab reports it matches Inkling's per-task results on terminal-based coding work at roughly a fifth of the cost. Available on Together AI.

‍

Total Parameters (12B active)

276B

MoE routing each token to 6 of 256 experts plus 2 shared experts

Context Window

Entire repositories or thousands of pages in a single request

Audio & Vision Input

Native

Speech, images, and text processed jointly by the same model

Model key capabilities

Multimodal Input: Accepts text, images, and audio in a single request, processed jointly rather than through separate pipelines
Variable Thinking Effort: Adjusts reasoning depth per request to balance cost against quality
Agentic Coding: Handles real-repository software engineering tasks and multi-tool agent workflows at a compact active parameter count
Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure

Performance benchmarks

Model	GPQA Diamond	HLE	SWE-bench verified
Inkling Small	89.5	31.6	80.2
Related open-source models
Competitor closed-source models
Claude Fable 5	92.6%	53%
Claude Opus 5	93.2%	53%
GPT-5.6 Sol	94.1%	47%
Grok 4.5	93.1%	40%

API usage

cURL
Python
Typescript

Endpoint:

thinkingmachines/Inkling-Small

curl -X POST "https://api.together.xyz/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "thinkingmachines/Inkling-Small",
    "messages": [
      {
        "role": "user",
        "content": "What are some fun things to do in New York?"
      }
    ]
}'

from together import Together

client = Together()

response = client.chat.completions.create(
  model="thinkingmachines/Inkling-Small",
  messages=[
    {
      "role": "user",
      "content": "What are some fun things to do in New York?"
    }
  ]
)
print(response.choices[0].message.content)

import Together from 'together-ai';
const together = new Together();

const completion = await together.chat.completions.create({
  model: 'thinkingmachines/Inkling-Small',
  messages: [
    {
      role: 'user',
      content: 'What are some fun things to do in New York?'
     }
  ],
});

console.log(completion.choices[0].message.content);

Model card
Architecture Overview:
• 42-layer decoder-only transformer with a sparse Mixture-of-Experts feed-forward backbone
• Each token routes to 6 of 256 experts, plus 2 shared experts active on every token
• Hybrid local and global attention layers with a context window of up to 1M tokens
• Natively multimodal: images encoded via a hierarchical patch encoder and audio via discrete token encoding, projected into a shared hidden space

Training Methodology:
• Trained with the same recipe as Inkling, Thinking Machines Lab's flagship open-weights model, at roughly a quarter of its total parameter count
• Built for breadth across domains, intended to make downstream fine-tuning effective from smaller amounts of data

Performance Characteristics:
• Real-repository coding (SWE-bench Verified, resolving actual GitHub issues): 80.2%; SWE-bench Pro (Public): 55.9%
• Terminal coding (Terminal-Bench 2.1, completing tasks in a command-line environment): 64.69
• Expert knowledge and science (GPQA Diamond, graduate-level science questions): 89.5%; Humanity's Last Exam (text only): 31.6%, or 47.8% multimodal with tools
• Math competitions (AIME 2026 / HMMT Feb 2026): 95.5% / 90.2%
• Abstract reasoning (ARC-AGI-1 / ARC-AGI-2, novel visual puzzle solving): 84.0% / 40.1%
• Tool orchestration (MCP Atlas, agent tasks across many tools): 79.2%; web research (BrowseComp with context management): 77.4%
• Vision (MMMU Pro university-level multimodal reasoning: 74.0%; CharXiv RQ chart reasoning: 77.4%, or 82.3% with Python)
• Audio (VoiceBench spoken-input assistant tasks: 90.1%; MMAU audio understanding: 77.0%)
• Instruction following (IFBench): 82.2%; multilingual knowledge (Global-MMLU-Lite): 86.7%
• Thinking Machines Lab reports identical per-task Terminal-Bench 2.1 performance to Inkling at about one-fifth of the cost
‍
Prompting
Together AI API Access:
• Access Inkling-Small via Together AI APIs using the endpoint thinkingmachines/Inkling-Small
• Authenticate using your Together AI API key in request headers
• Supports native function calling, with text, image, and audio input in a single request
• Available on Together AI serverless and dedicated infrastructure
‍
Applications & use cases
Cost-Efficient Agentic Coding:
• Resolve repository issues and run terminal-based tasks with a 12B-active-parameter cost profile
• Scale concurrent coding agents where per-request cost matters as much as capability
• Dial thinking effort down for routine tasks and up for hard ones within the same deployment

Voice & Multimodal Assistants:
• Process spoken input directly alongside text and images without a separate transcription step
• Build assistants that reason over screenshots, documents, and audio context in one call
• Serve high-volume multimodal traffic at small-model inference cost

Long-Context Workflows:
• Hold large mixed-media context, up to 1M tokens, across extended sessions
• Cross-reference big document sets and codebases in a single request
• Run sustained multi-turn agent loops without losing earlier context
‍

Related models

Model specifications

Model data

Model provider
Thinking Machine Labs
Type
Chat
Reasoning
Vision
Audio
Features
Function Calling
Deployment
Serverless
Endpoint
thinkingmachines/Inkling-Small
Parameters
276B
Activated parameters
12B
Context length
1M
Input price
$0.50 / 1M tokens
$0.10 (cached)/1M
Output price
$1.20 / 1M tokens
Input modalities
Text
Image
Audio
Output modalities
Text

Released
July 29, 2026
Quantization level
FP8
Category
Chat

Run in Playground

Quickstart docs

Deploy model

Inkling Small

About model

API usage

Model card

Prompting

Applications & use cases