Back to list
Product LaunchAI HardwareOpen SourceDeep Learning

Tiny Corp Unveils Tinybox: High-Performance Offline AI Hardware Supporting Massive Parameter Models

Tiny Corp has officially launched the tinybox, a specialized computer designed to run powerful neural networks offline. Built on the tinygrad framework, which simplifies complex networks into three fundamental operation types (ElementwiseOps, ReduceOps, and MovementOps), the tinybox is available in multiple configurations including 'red', 'green', and the upcoming 'exa' scale. The top-tier 'green v2' model boasts 3086 TFLOPS of FP16 performance and 384 GB of GPU RAM, while the ambitious 'exabox' aims for exascale performance. Tiny Corp is currently leveraging its funded status to expand its team of software, hardware, and operations engineers, prioritizing contributors to the tinygrad open-source ecosystem.

Hacker News

Key Takeaways

  • Hardware Availability: The tinybox is now shipping in 'red' and 'green' versions, with a high-end 'exabox' in development.
  • Performance Specs: The 'green v2' model features 4x RTX PRO 6000 GPUs, delivering 3086 TFLOPS and 384 GB of GPU RAM.
  • Software Foundation: Powered by tinygrad, a framework that reduces complex neural networks to Elementwise, Reduce, and Movement operations.
  • Expansion: Tiny Corp is actively hiring full-time engineers and interns, specifically seeking those who have contributed to the tinygrad codebase.

In-Depth Analysis

The tinygrad Framework Philosophy

At the heart of the tinybox is tinygrad, positioned as the fastest-growing neural network framework. Its architectural philosophy centers on extreme simplicity. Rather than maintaining a massive library of disparate operations, tinygrad breaks down complex networks into three core OpTypes: ElementwiseOps (Unary, Binary, and Ternary operations like SQRT and ADD), ReduceOps (such as SUM and MAX), and MovementOps. The latter are virtual operations that utilize a ShapeTracker for copy-free data manipulation, including RESHAPE and PERMUTE. This streamlined approach aims to solve the mystery of traditional CONVs and MATMULs through code efficiency rather than abstraction bloat.

Hardware Tiers: Red, Green, and Exa

Tiny Corp offers distinct hardware paths to cater to different computational needs. The 'red v2' serves as an entry point with 4x 9070XT GPUs and 778 TFLOPS of performance. The 'green v2' significantly scales up capabilities using 4x RTX PRO 6000 GPUs, providing 384 GB of GPU RAM and a massive 3086 TFLOPS. For extreme scale, the 'exabox' is detailed with 720x RDNA5 AT0 XL GPUs, aiming for ~1 EXAFLOP of performance and over 25,000 GB of GPU RAM. These systems are designed for high-bandwidth tasks, with the green model utilizing PCIe 5.0 x16 and the exabox featuring 400 GbE networking.

Operational and Recruitment Strategy

Tiny Corp is transitioning from a development project to a funded commercial entity. Their recruitment strategy is uniquely tied to their open-source roots; candidates for software, hardware, and operations roles are generally not considered unless they have already contributed to the tinygrad framework. This "bounty-to-hire" pipeline allows the company to judge fit through practical contributions while paying developers for their work.

Industry Impact

The introduction of the tinybox represents a shift toward accessible, high-performance offline AI compute. By combining a simplified software stack (tinygrad) with powerful consumer and professional-grade GPUs, Tiny Corp provides an alternative to cloud-dependent AI development. The focus on "copy-free" movement operations and a reduced set of OpTypes suggests a push for higher efficiency in how hardware resources are utilized, potentially lowering the barrier for running large-scale models with billions of parameters locally.

Frequently Asked Questions

Question: What are the primary differences between the red and green tinybox models?

The red v2 uses 4x 9070XT GPUs with 64 GB of GPU RAM, while the green v2 utilizes 4x RTX PRO 6000 GPUs with 384 GB of GPU RAM and significantly higher TFLOPS (3086 vs 778).

Question: How does tinygrad handle complex operations like convolutions?

tinygrad simplifies all complex networks into three basic types: ElementwiseOps, ReduceOps, and MovementOps. It avoids traditional bulky implementations of CONVs by breaking them down into these fundamental operations.

Question: Can I apply for a job at Tiny Corp without prior experience with their framework?

According to the company, applications for software, hardware, and operations roles will not be considered unless the applicant has already contributed to the tinygrad framework.

Related News

Google’s Gemini Decision 1 Picks Answers Instead of Writing Them Across Five Input Types
Product Launch

Google’s Gemini Decision 1 Picks Answers Instead of Writing Them Across Five Input Types

Google’s Gemini Decision 1 is a model designed to pick answers rather than write them, according to Shittu Olumide’s report for KDnuggets. It reads text, images, video, audio, and PDFs in a single forward pass and supports up to 262,144 input tokens per request. The central distinction is between what the model reads and what it does with that input: it accepts several kinds of material, but its stated task is answer selection, not answer writing. The token figure describes the maximum input per request, not how much the model outputs. The report does not explain how answers are presented for selection or provide results showing how well the model performs.

NVIDIA RTX Spark Compact Desktops Now Available From ASUS, Dell and MSI for Local AI
Product Launch

NVIDIA RTX Spark Compact Desktops Now Available From ASUS, Dell and MSI for Local AI

NVIDIA announced on November 18, 2026, that RTX Spark compact desktops from ASUS, Dell and MSI were available. The systems bring local AI agents, image and video generation, and model fine-tuning to a dedicated Windows PC. NVIDIA lists up to 128GB of unified memory and its AI software stack as features that give larger local models more room, alongside RTX-accelerated creative and development work. The launch centers on bringing these different workloads together in a compact desktop rather than presenting the PC as a system for a single AI task. The available excerpt does not include pricing, partner-specific configurations or performance benchmarks, so the announcement establishes what the systems are designed to support but does not show how the individual desktops compare.

Instagram Rolls Out Content Protection Feature to Help Creators Detect and Review Copycat Videos Across Platform
Product Launch

Instagram Rolls Out Content Protection Feature to Help Creators Detect and Review Copycat Videos Across Platform

Instagram has begun rolling out a new content protection tool aimed at supporting video creators by detecting unauthorized copies of their work. According to reporting from The Verge, the feature functions as an automated detection system designed specifically for eligible creators who enroll in the program. Once activated, the system actively searches for visual matches of original video uploads and notifies creators when matching content is identified across Instagram. This update empowers enrolled creators by giving them the visibility to review duplicated video assets directly within the platform, offering a formal mechanism to track copies of their original multimedia work.