Inkling 2 Small

Long-horizon agentic coding, visual creation, and knowledge work with 1M context

Try now

Read Docs

About model

Inkling 2 Small has 594B parameters, 17B active parameters and a 1M-token context window. On Together AI it costs $0.30 per 1M input tokens and $1.00 per 1M output tokens.

Inkling 2 Small is the first model in Thinking Machines Lab's second Inkling generation: a Mixture-of-Experts model with 594B total parameters and 17B active per token, native image input, a 1M-token context window, and variable reasoning effort. It is more capable across the board than the first generation, particularly in long-horizon autonomous work, with gains in agentic coding, vision, knowledge work, and reasoning that come from improvements at every stage of training, including more efficient pre-training run natively in NVFP4 and a single large, unified reinforcement learning run. The model uses vision in the loop to interpret and refine the websites, games, and 3D designs it builds, handles finance and legal work over spreadsheets, filings, and documents, and is designed to respond well to fine-tuning, learning faster than Inkling Small on the same data. Released under Apache 2.0. Available on Together AI.

‍

Total Parameters (17B active)

594B

Sparse MoE routing each token to 8 of 384 experts

Context Window

Long agent trajectories, large repositories, and document sets

Image Input

Native

Vision in the loop to check and refine the interfaces it builds

Model key capabilities

Agentic Coding & Autoresearch: Trained on hundreds of thousands of long-horizon coding tasks across agent harnesses, following instructions closely and verifying its own work, plus autonomous ML and kernel-optimization research
Visual Interpretation & Creation: OCR, grounding, infographic reading, and visual reasoning, with visual feedback used to iteratively build websites, games, and 3D designs
Agentic Knowledge Work: Looks up information, calls tools over MCP, and produces deliverables, with particular strength in finance and quantitative reasoning over spreadsheets and filings
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 2 Small		93.0						25.4	72.1
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-2-Small

curl -X POST "https://api.together.ai/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "thinkingmachines/Inkling-2-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-2-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-2-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:
• 48-layer decoder-only transformer with a sparse Mixture-of-Experts backbone: 594B total parameters, 17B active, each token routed to 8 of 384 experts
• Hybrid attention combining local and global layers, with a 1M-token context window
• Natively multimodal: a 292M parameter vision encoder based on Perception Encoder uses 20x20-pixel patches with 2x2 spatial pooling, interleaving image patches with text tokens in a shared hidden space
• Supports MXFP8 and NVFP4 numerics; image inputs between 40 and 4,096 pixels per dimension

Training Methodology:
• Initial pre-training on 8.4T tokens of natural language, code, and image data, run natively in NVFP4, followed by 3.8T tokens of higher-quality, newer data at 512K context
• Further phases of reasoning and agent trajectories at 1M context, about 320B and 34B tokens, filtered for correctness and alignment
• A single joint reinforcement learning run of 1,075 steps and more than 17 million rollouts over 9 days on 2,048 GPUs, spanning coding, knowledge work, STEM, web and app building, vision, tool use, search, and safety, with a grader tailored to each task family
• Effort levels sampled during RL with a length penalty, training the model to trade off speed against thoroughness
• Alignment work across the stack: filtering misbehavior from midtraining data, a general alignment reward in RL, environments that reward giving up on impossible tasks, and on-policy self-distillation from a teacher prompted with a conduct spec

Performance Characteristics:
• Results reported by Thinking Machines Lab at effort 0.99 (xhigh) and temperature 1.0, using externally reported evaluations where available
• Artificial Analysis Intelligence Index: 30.6 (v5.0), far above earlier Inkling models
• Coding (DeepSWE v1.1: 72.1%; SWE-bench Verified: 82.2%; SWE-bench Pro v2 Hard: 86.3%; FrontierCode v2.0: 25.4%; Terminal-Bench 2.1 with Pi harness: 85.1%; Vibe Code Bench v1.1: 81.6%)
• Knowledge work (GDPval-AA v2.1: 1630; AA-Briefcase v1.1: 1527; JobBench: 70.5%; Finance Agent v2.1: 67.5%; Vals Legal Index: 73.9%; MCP Atlas: 85.4%)
• Vision (MMMU Pro: 81.6%; ScreenSpot Pro: 83.8%, or 91.5% with Python; CharXiv RQ: 85.6%; MathVision: 92.8%; PerceptionBench: 64.7%)
• Reasoning and long context (GPQA Diamond: 93.0%; HMMT February 2026: 95.8%; AA-LCR v1.1: 83.3%; LongBench v2: 70.6%)
• Thinking Machines reports the top DesignArena Website score among open-weight models in its November 5 leaderboard snapshot
• Alignment: misbehavior rate on internal evaluations of 14.9, down from 66.8 for Inkling and 75.2 for Inkling Small
‍
Prompting
Together AI API Access:
• Access Inkling 2 Small via Together AI APIs using the endpoint thinkingmachines/Inkling-2-Small
• Authenticate using your Together AI API key in request headers
• Reasoning effort is variable per request; published results use effort 0.99 (xhigh) with temperature 1.0
• Supports text and image input across a 1M-token context window
• Available on Together AI serverless and dedicated infrastructure
‍
Applications & use cases
Agentic Software Engineering:
• Run long-horizon coding agents that follow instructions closely and verify their own changes
• Optimize GPU kernels and run autonomous ML experiments
• Hold large repositories and long trajectories in the 1M-token window

Visual Building & Design:
• Build websites, games, and 3D designs that the model inspects and refines visually
• Read screenshots, charts, and infographics as part of agent workflows
• Ground UI tasks in on-screen element locations

Finance & Legal Work:
• Analyze spreadsheets, filings, and documents for financial research and modeling
• Produce professional deliverables through tool use over MCP
• Support legal research and document review

Custom Models:
• Use as a base for fine-tuning on domain data, where it learns faster than earlier Inkling models
• Tune reasoning effort per request to balance speed and thoroughness
• Deploy specialized variants for enterprise workflows
‍