> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Thinking Machines Shrinks Its AI to 25% Size With Near-Identical Performance
- URL: https://wire.fourthweb.ai/thinking-machines-shrinks-its-ai-to-25-size-with-near-identical-performance/
- Published: 2026-07-31T00:14:22.000Z
- Updated: 2026-07-31T05:32:42.000Z
- Description: The era of "bigger is better" AI just hit a wall, and the timing couldn't be more convenient for companies stuck on waitlists for H100s.
- Author: Travis Wright
- Tags: AI Agent Economy, AI Infrastructure, OpenAI, Anthropic, Nvidia

**The era of "bigger is better" AI just hit a wall, and the timing couldn't be more convenient for companies stuck on waitlists for H100s.**

### The Summary

- [Thinking Machines released Inkling-Small, a 276B-parameter open source model that matches 99% of its 975B-parameter predecessor's performance](https://venturebeat.com/technology/thinking-machines-debuts-inkling-small-open-source-ai-model-nearing-performance-of-predecessor-at-about-1-4-size?ref=wire.fourthweb.ai) while using one-quarter the size
- Uses 12B active parameters per token vs. 41B in the original, cutting [compute](https://wire.fourthweb.ai/tag/ai-infrastructure/) costs and deployment requirements by roughly 70%
- Released under Apache 2.0 license with full weights on Hugging Face, making it immediately forkable and commercially deployable
- Launch pricing at $0.58 per million input tokens represents a direct challenge to closed-model economics

### The Signal

Two weeks. That's how long it took Mira Murati's new company to shrink a flagship model by 75% without meaningful performance loss. Either Thinking Machines deliberately released an inefficient first model to establish a baseline, or they just proved that most of the AI industry has been burning money on parameters that don't materially improve output.

The numbers tell the efficiency story. [Inkling-Small scores within one point of the original Inkling on the Artificial Analysis Intelligence Index](https://venturebeat.com/technology/thinking-machines-debuts-inkling-small-open-source-ai-model-nearing-performance-of-predecessor-at-about-1-4-size?ref=wire.fourthweb.ai), despite using 12 billion active parameters per token instead of 41 billion. That's not a marginal improvement. That's a fundamental rethinking of what parameters actually do work.

> "The appeal is not simply that Inkling-Small is smaller. It is that developers appear to give up relatively little capability while reducing the model's compute requirements."

This matters because the [GPU](https://wire.fourthweb.ai/tag/compute-wars/) shortage isn't really over. It's just evolved from "impossible to get" to "extremely expensive to run at scale." A model that delivers 99% of the capability at 25% of the compute cost changes the economics for every company that's been waiting for inference prices to drop before committing to AI-native workflows.

The multimodal piece is where this gets interesting for actual deployment:

- Text, image, and audio inputs with text output
- One million token context window
- Full fine-tuning support through Tinker API
- Still too large for consumer hardware but viable for mid-sized enterprise GPU clusters

What Thinking Machines is doing here is creating the open source equivalent of what [Anthropic](https://wire.fourthweb.ai/tag/anthropic/) did with Claude Sonnet versus Opus. Give developers a model that's "good enough" for 90% of use cases, make it dramatically cheaper to run, and watch adoption accelerate because the cost-benefit calculation finally makes sense.

The Apache 2.0 license is the real power move. No usage restrictions. No revenue sharing. No mysterious fine print about derivative works. Companies can take these weights, fine-tune them on proprietary data, and deploy without looking over their shoulders. That's the license you use when you want maximum distribution and you're confident your moat is speed of iteration, not legal lock-in.

### The Implication

Watch what happens to closed-model pricing over the next quarter. When an open source model with full commercial rights delivers this much capability at these price points, the premium for proprietary models shrinks fast. Companies that were waiting for "good enough and cheap enough" just got their signal.

For developers, the play is clear: start testing Inkling-Small against whatever you're running now. If it passes your evals, you just cut your inference bill by 60-70% and gained the ability to self-host if you ever want to. That's not a marginal improvement. That's a different game.

### Sources

[VentureBeat](https://venturebeat.com/technology/thinking-machines-debuts-inkling-small-open-source-ai-model-nearing-performance-of-predecessor-at-about-1-4-size?ref=wire.fourthweb.ai)