Inkling

Multimodal reasoning model with variable thinking effort and native audio and vision

Try now

read docs

About model

Inkling is Thinking Machines Lab's first open-weights model, a Mixture-of-Experts transformer with 975B total parameters and 41B active, natively trained to reason over text, images, and audio in a shared hidden space. Each token routes to 6 of 256 experts plus 2 shared experts, with hybrid local and global attention supporting a 524K-token context window. The model varies its thinking effort to balance cost against performance and is built to give calibrated answers rather than confident wrong ones, with 87.9% on GPQA Diamond and 77.6% on SWE-bench Verified among its reported results. Released under Apache 2.0. Available on Together AI.

‍

Total Parameters (41B active)

975B

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

Context Window

524K

Hybrid local and global attention for long multimodal sessions

SWE-bench Verified

77.6%

Agentic coding performance on real-world software engineering tasks

Model key capabilities

Native Multimodal Reasoning: Joint understanding of text, image, and audio input in a shared hidden space, with 87.9% on GPQA Diamond
Variable Thinking Effort: Balances cost against performance by adjusting reasoning depth per request
Calibrated Responses: Trained to give calibrated answers rather than confident wrong ones
Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure

Performance benchmarks

Model	FrontierMath Tier 4	GPQA Diamond	HLE	SciCode	GDPval-AA	Terminal-Bench 2.1	Agent Arena	FrontierCode	DeepSWE
Inkling		87.2%	30%	46%		55%	-9.7pp	14.0%
Related open-source models
Competitor closed-source models
Claude Fable 5	87.8%	92.6%	53%	60%	62%	85%		53.5%	70%
Claude Opus 5	73.2%	93.2%	53%	56%	68%	89%	+16.4pp	53.4%	74%
GPT-5.6 Sol	82.9%	94.1%	47%	56%	61%	88%		47.5%	73%
Grok 4.5	24.4%	93.1%	40%	54%	51%	82%	+4.2pp	42.4%	54%
GPT-5.6 Luna	61.0%	91.1%	37%	53%	54%	81%		39.8%	67%

API usage

cURL
Python
Typescript

Endpoint:

thinkingmachines/Inkling

curl -X POST https://api.together.xyz/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -d '{
    "model": "thinkingmachines/Inkling",
    "messages": [{
      "role": "user",
      "content": "Given two binary strings `a` and `b`, return their sum as a binary string"
    }]
  }'

from together import Together

client = Together()
response = client.chat.completions.create(
  model="thinkingmachines/Inkling",
  messages=[
  	{
	    "role": "user", 
      "content": "Given two binary strings `a` and `b`, return their sum as a binary string"
    }
 ],
)

print(response.choices[0].message.content)

import Together from "together-ai";

const together = new Together();

async function main() {
  const response = await together.chat.completions.create({
    model: "thinkingmachines/Inkling",
    messages: [{
      role: "user",
      content: "Given two binary strings `a` and `b`, return their sum as a binary string"
    }]
  });
  
  console.log(response.choices[0]?.message?.content);
}

main();

Model card
Architecture Overview:
• 66-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 524K-token context window
• Natively multimodal: images encoded via a hierarchical patch encoder and audio via discrete token encoding, with all modalities projected into a shared hidden space

Training Methodology:
• Trained from scratch on a broad mix of text, image, audio, and video content drawn from public sources, third-party data, and synthetic generation
• Training data curation included deduplication and filtering to remove low-quality content and advance safety objectives
• Trained for breadth across domains rather than a single vertical, intended to make downstream fine-tuning more effective from a smaller amount of data

Performance Characteristics:
• Reasoning: 87.9% GPQA Diamond, 97.1% AIME 2026, 30.0% Humanity's Last Exam (text only) and 46.0% with tool use
• Agentic coding: 77.6% SWE-bench Verified, 54.3% SWE-bench Pro (Public), 62.5 Terminal-Bench 2.1
• Agentic (general): 74.1% MCP Atlas, 78.0% BrowseComp (with context management)
• Chat and factuality: 88.7% Global MMLU, 43.9% SimpleQA Verified
• Vision: 73.3% MMMU Pro (Standard 10), 78.1% ChartQA RQ
• Audio: 58.8% Audio MC, 77.2% MMAU, 91.4% VoiceBench
• Safety: 99.1% StrongReject, 79.42 FORTRESS Harmful (public), 95.51 FORTRESS Benign (public)
‍
Prompting
Together AI API Access:
• Access Inkling via Together AI APIs using the endpoint thinkingmachines/inkling
• Authenticate using your Together AI API key in request headers
• Supports native function calling and JSON mode for structured outputs
• Accepts text, image, and audio input in a single request
• Available on Together AI serverless and dedicated infrastructure
‍
Applications & use cases
Multimodal Agentic Systems:
• Build agents that reason jointly over text, images, and audio in one call
• Combine native function calling with JSON mode for structured, tool-driven workflows
• Run agentic coding tasks with SWE-bench-level performance through the Together endpoint

Voice & Audio Applications:
• Process spoken input directly alongside text and images without a separate transcription step
• Build voice-aware assistants that reason over audio context and respond with calibrated answers
• Combine audio understanding with long-context history across extended sessions

Long-Context Multimodal Workflows:
• Hold large mixed-media context — documents, images, and audio — across a 524K-token window
• Run sustained multi-turn agent sessions without losing earlier multimodal context
• Analyze and cross-reference multimodal material in a single extended session
‍

Related models

Model specifications

Model data

Model provider
Thinking Machine Labs
Type
Chat
Reasoning
Vision
Audio
Features
Function Calling
JSON Mode
Deployment
Serverless
Endpoint
thinkingmachines/Inkling
Parameters
975B
Activated parameters
41B
Context length
524K
Input price
$1.00 / 1M tokens
$0.17 (cached)/1M
Output price
$4.05 / 1M tokens
Input modalities
Text
Image
Audio
Output modalities
Text

Released
July 15, 2026
Quantization level
NVFP4
Category
Code

Run in Playground

Quickstart docs

Deploy model

Inkling

About model

API usage

Model card

Prompting

Applications & use cases