Skip to content
Frontier Leaf FL markFLFrontierLeaf

A field journal of new AI developments.

Model updates

Inkling 2 Small automates AI training tasks with vision verification

Thinking Machines releases a model that doubles forecasting scores and checks code visually, amid reports of a major funding round.

Depth
68 hoursInkling 2 Small pre-trained in 68 hours on GB300 NVL72 racks · reported
4BThe model doubled Qwen3.5-4B's weather forecasting score in 4.5 hours · reported
4.5 hoursThe model doubled Qwen3.5-4B's weather forecasting score in 4.5 hours · reported
1 outlet · newest source 378h ago
Inkling 2 Small automates AI training tasks with vision verification

Key points

  • Inkling 2 Small pre-trained in 68 hours on GB300 NVL72 racks.
  • The model doubled Qwen3.5-4B's weather forecasting score in 4.5 hours.
  • A vision encoder allows visual inspection of generated e-commerce pages.

Thinking Machines Labs has released Inkling 2 Small, a large language model designed to automate artificial intelligence training tasks. Inkling 2 Small launched late Thursday as Business Insider reported that Thinking Machines is seeking a $5 billion funding round led by Accel.This new AI model can automatically improve other AI systems and check its own work visually.Inkling 2 Small demonstrates automated RLHF and visual verification capabilities on GB300 hardware.

Evidence: reportedThinking Machines debuts Inkling 2 Small amid reported $5B fundraising push

aifleet-siliconangle

Evidence: reportedThinking Machines debuts Inkling 2 Small amid reported $5B fundraising push

aifleet-siliconangle

The facts

Thinking Machines finished the initial pre-training in 68 hours, then refined the model through a nine-day reinforcement learning process that included several intermediate steps.

Evidence: reportedThinking Machines debuts Inkling 2 Small amid reported $5B fundraising push

aifleet-siliconangle

Thinking Machines improved Inkling 2 Small’s coding skills by having it complete hundreds of thousands of long-horizon programming challenges. These included complex "autoresearch" exercises requiring the model to post-train neural networks, optimise LLM kernels and perform other AI research work.

Evidence: reportedThinking Machines debuts Inkling 2 Small amid reported $5B fundraising push

aifleet-siliconangle

The facts

Thinking Machines evaluated Inkling 2 Small's machine learning capabilities by having it fine-tune Qwen3.5-4B. According to the company, the model doubled the open-source LLM's forecasting score in just 4.5 hours.

Evidence: reportedThinking Machines debuts Inkling 2 Small amid reported $5B fundraising push

aifleet-siliconangle

Evidence: company claimThinking Machines debuts Inkling 2 Small amid reported $5B fundraising push

aifleet-siliconangle

The facts

The LLM is also adept at automating business tasks. The model surpassed Spark Open, Kimi K3, and various other open-source systems on three widely used knowledge work benchmarks and achieved superior results on several financial task datasets.

Evidence: reportedThinking Machines debuts Inkling 2 Small amid reported $5B fundraising push

aifleet-siliconangle

Building on this

Inkling 2 Small features a vision encoder capable of image analysis. This allows the model to visually assess the quality of its own generated images. When tasked with building an online shop, the model can scan the code for errors and visually examine product pages to ensure they display properly. The model contains 594 billion parameters, 292 million of which belong to the vision encoder, while approximately 17 billion activate with each user prompt.

Evidence: reportedThinking Machines debuts Inkling 2 Small amid reported $5B fundraising push

aifleet-siliconangle

Evidence: reportedThinking Machines debuts Inkling 2 Small amid reported $5B fundraising push

aifleet-siliconangle

Why it matters

  • The ability to visually inspect generated web pages reduces the need for manual QA in automated coding workflows.
  • Rapid fine-tuning of open-source models like Qwen3.5-4B suggests faster iteration cycles for custom AI applications.
  • High performance on financial and knowledge work benchmarks indicates potential for enterprise automation tasks.
  • The reported $5 billion fundraising round may influence future pricing and availability of the model.

Topics: models, AI training, vision encoders, Thinking Machines, large language models

Updated

Then / now / next

Then

Inkling 2 Small was not publicly available.

Now

The model is released and can automate AI training and business tasks.

Next

Watch for updates on the $5 billion fundraising round led by Accel.

Story timeline

  1. Then Inkling 2 Small was not publicly available.
  2. 06 Nov 2026 Thinking Machines debuts Inkling 2 Small amid reported $5B fundraising push
  3. Now The model is released and can automate AI training and business tasks.
  4. Next Watch for updates on the $5 billion fundraising round led by Accel.

For readers

Developers

  • Can use vision encoder to visually verify generated code and web pages.
  • Can automate fine-tuning of open-source models like Qwen3.5-4B.

Investors

  • Thinking Machines is reportedly raising a $5 billion round led by Accel.
  • The company expects a $40 billion valuation.

Researchers

  • Model trained on complex autoresearch exercises including neural network post-training.

Numbers in context

68 hoursreported4.5 hoursreportedCircle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.

Ask the article

Local keyword search of this article; no generated answers.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Questions readers ask

How long did it take to pre-train Inkling 2 Small?

The initial pre-training phase was completed in 68 hours.

What hardware was used for training?

Thinking Machines used a cluster of GB300 NVL72 racks with Blackwell Ultra graphics cards.

Can the model handle visual tasks?

The model features a vision encoder capable of analyzing images and visually verifying the quality of its generated graphics.

Connected stories

This storyDartmouth provost resigns over undisclosed AI use in academic writingDartmouth provost resigns over uDevelopers face Feb 2027 sunset for GPT-5.5-cyber-preview as new GPT-6-Cyber requires Daybreak setupDevelopers face Feb 2027 sunset OpenAI: Fair Use Law Powers AI Progress & Helps Disabled People Access ContentOpenAI: Fair Use Law Powers AI PDevelopers run frontier 145B coding AI on laptops, cutting cloud costs and keeping data privateDevelopers run frontier 145B cod

Reading receipt

  • Inkling 2 Small pre-trained in 68 hours on GB300 NVL72 racks.
  • The model doubled Qwen3.5-4B's weather forecasting score in 4.5 hours.
  • A vision encoder allows visual inspection of generated e-commerce pages.

Still open: What are the specific costs or access requirements for using Inkling 2 Small?

Still open: How does the model perform on tasks outside of coding, finance and knowledge work?

Time spent: 0:00 · Depth: Full

Sources

  1. SiliconANGLE AI

Related stories

Background reading