America Finally Has a Frontier Open-Weights Model. Thinking Machines Just Changed the Game.

For two years, if you wanted an open-weights AI model that didn't suck, your options were Chinese. GLM-5.2. DeepSeek V4. Qwen3. Kimi K2.7. All excellent. All from labs in Beijing, Hangzhou, and Shenzhen.

The American side? Meta fumbled Llama 4 so badly that HN commenters are writing its obituary. Google's Gemma line is solid but lives in a different weight class. The frontier labs — OpenAI, Anthropic — aren't releasing weights at all.

Then yesterday, Thinking Machines Lab dropped Inkling. And the landscape just shifted.

What Inkling Actually Is

Thinking Machines Lab — the company Mira Murati built after leaving OpenAI — released their first model. It's called Inkling, and the specs are serious:

• 975B total parameters, 41B active (Mixture-of-Experts)
• 1 million token context window
• Trained on 45 trillion tokens of text, images, audio, and video
• Native multimodal — reasons over text, vision, and audio in the same model
• Controllable thinking effort — you can dial reasoning depth from 0.2 to 0.99
• Fully open weights on HuggingFace

They're also previewing Inkling-Small, a lighter version with 12B active parameters trained with the same recipe. Same multimodal capabilities, lower cost, lower latency.

This isn't a toy. It's a frontier-class model with full weights available for anyone to download, fine-tune, or deploy. The training data — 45 trillion tokens spanning four modalities — puts it in the same league as the models that have been dominating leaderboards for months.

The "Finally, An American Open Model" Story

Here's the part that has the AI community buzzing.

Go to any thread about open-weights models and you'll find the same conversation. Someone asks which open model to use. The answers are always the same: GLM, DeepSeek, Qwen, Kimi. Someone points out these are all Chinese. Someone else says "yeah, and?" Because until now, American open models haven't been competitive at the frontier.

The top comment on Hacker News yesterday put it bluntly: "This is the first competitive non-Chinese open weights model since what, Llama 3?"

Another user: "America needs its own DeepSeek or Z.ai. A lot of people root for open Chinese models to win because they have no other choice. Thinking Machines might be it."

That "no other choice" line stings. The world's AI developers have been relying on Chinese labs for open weights because American companies either went closed (OpenAI, Anthropic) or went sideways (Meta's Llama 4 benchmark controversy, then the quiet shift toward proprietary releases like Muse Spark). The void was real.

Inkling doesn't solve every problem. Thinking Machines themselves are upfront about this — they say it's not the strongest model available, open or closed. But it's competitive across the board: agentic coding, reasoning, instruction following, vision, and audio. It's a generalist, not a specialist optimized to game one benchmark.

And it's American. For organizations with data sovereignty requirements, export control concerns, or just geopolitical unease about depending on infrastructure controlled by a strategic competitor, that matters more than people like to admit out loud.

The Audio Angle Nobody's Talking About

Here's something buried in the specs that deserves more attention: Inkling is the largest open-weights model that supports audio.

Not "audio via a separate speech-to-text pipeline bolted on afterward." Native audio reasoning. It scored 77.2% on MMAU (Multi-modal Audio Understanding) and 56.6% on Audio MC. The only model that beats it on audio is Gemini 3.1 Pro — which is closed-source and costs money per API call.

If you're building voice agents — and a lot of people are, given that OpenAI's GPT-Live just made real-time voice AI feel real — having an open model that natively understands audio is a big deal. You're not chaining Whisper → LLM → TTS anymore, with latency stacking up at each hop. The model hears, thinks, and responds in one pass.

This is also strategic. Thinking Machines recently introduced their "interaction models" concept — AI systems designed for real-time voice and vision collaboration between humans and machines. Inkling was built to be the reasoning engine behind that vision. They're not just releasing a model. They're releasing the foundation for a product roadmap that goes way beyond text chat.

Controllable Thinking: The Feature That Actually Matters

Every reasoning model has the same problem: sometimes you need it to think hard, and sometimes you need it to just answer. GPT and Claude handle this with separate "thinking" tiers or model variants. You pick "high reasoning" or "fast mode" and hope you chose right.

Inkling does something cleaner.

You set a single parameter — thinking effort, from 0.2 to 0.99 — and the model adjusts how much internal reasoning it does before responding. The benchmark chart on their announcement page shows Inkling matching Nemotron 3 Ultra on Terminal Bench 2.1 (agentic coding) at roughly one-third the tokens.

One-third. That's not a marginal optimization. That's the difference between a model that costs $3 per task and one that costs $1. Multiply that across thousands of agent runs per day and it's the difference between a product that's economically viable and one that burns money.

For anyone building agents that make dozens of tool calls per workflow, this is the feature to watch. Most tasks don't need maximum reasoning depth. A simple file lookup doesn't need 10,000 tokens of chain-of-thought. Being able to dial it down without swapping models or losing capability is a genuine architectural advantage, and it's something no other open model offers at this level of control.

The Self-Fine-Tuning Demo

Thinking Machines did something clever to show off their fine-tuning platform, Tinker. They had Inkling fine-tune itself.

The model wrote its own fine-tuning job, ran it on Tinker, evaluated the results against a rubric it also wrote, and produced an improved version of itself. It's a demo, not a production capability — but it's a demo that makes a point: fine-tuning should be accessible enough that the model itself can drive the process.

This matters because fine-tuning has always been the hardest part of working with open models. You can download GLM-5.2 or DeepSeek weights today, but adapting them to your specific use case requires GPU clusters, training expertise, and weeks of iteration. Most teams give up and just call the API.

Tinker is Thinking Machines' answer to that — a platform where you upload your data, pick a base model (Inkling, obviously), and get a fine-tuned model back without touching PyTorch. The open weights get you the model. Tinker gets you the customization. That's the business model — and unlike pure open-weights releases that generate goodwill but no revenue, this one has a clear path to actual money.

Can You Actually Run It Locally?

Let's be real. A 975B parameter model is not running on your laptop. It's not running on your desktop with a single 4090. This is a data center model — you need serious hardware or you're calling an API.

But the community moves fast. Unsloth already has GGUF and NVFP4 quantized versions on HuggingFace. There's a llama.cpp branch adding Inkling support. If you have the hardware — think multi-GPU setups with 200GB+ of VRAM — you can run it. The NVFP4 version compresses the model significantly while maintaining most of its quality. And llama.cpp support means CPU inference is technically possible, even if it's slow enough to make you question your life choices.

For everyone else, there's the Tinker playground, where you can chat with Inkling directly through a developer interface. It's not the same as local inference, but it lets you feel the model's capabilities before committing to anything.

Inkling-Small, with 12B active parameters, is the one to watch for practical local deployment. If it follows the pattern of other small MoE models — think Qwen 3.6 35B A3B, which runs on a single consumer GPU and punches well above its weight — it could be the model that brings Thinking Machines' multimodal recipe to local hardware. We don't have benchmarks yet, but the architecture is promising.

What This Means for the Open-Source AI Landscape

The open-weights leaderboard has been a one-region show for a while. China kept releasing bangers. America kept... not doing that.

Inkling changes the dynamic in three concrete ways:

1. The benchmark race just got more competitive. When the only serious open models come from one country, complacency sets in. GLM and DeepSeek are genuinely excellent, but competition drives excellence. An American lab entering the ring at this quality level pushes everyone to do better. The next DeepSeek or Qwen release now has to beat not just each other, but an American model with multimodal capabilities and a fine-tuning platform attached.

2. Fine-tuning becomes the product, not just the model. Thinking Machines isn't just releasing weights and walking away. They're building Tinker as the customization layer. If that works — if companies find it genuinely easier to fine-tune Inkling on Tinker than to DIY with LoRA adapters and rented A100s — it creates a template for how open-weights companies actually make money. Give away the model, charge for the platform. It's the open-source playbook that actually works: Red Hat did it with Linux. Thinking Machines might do it with AI.

3. Multimodal is now table stakes for open models. Inkling ships with native text, vision, and audio in one model. That raises the bar for everyone. The next Qwen or DeepSeek release won't be able to ship text-only and expect to compete. Users want models that can see, hear, and reason across modalities — and now there's an open option that does all three at scale.

The Honest Take

Inkling isn't going to dethrone GPT-5.6 or Claude Fable 5 tomorrow. Thinking Machines says so themselves, right on the announcement page: it's not the strongest model available, open or closed. It's a base model — strong, broad, multimodal, designed for customization rather than peak benchmark performance.

But that's exactly what the open-weights ecosystem needs right now. Not another model trying to win a leaderboard by 0.3 points. A genuinely useful foundation that developers can adapt, customize, and build products on top of without begging a closed lab for API access.

The fact that it's American is geopolitically significant. The fact that it's multimodal is technically significant. The fact that it comes with a real fine-tuning platform is commercially significant. And the fact that the weights are actually open — not "open" with a retrained-in-three-months kill switch — means you can build on it without existential risk to your product.

And the fact that Mira Murati — who was literally in the room when GPT was being built, who ran OpenAI's product organization through the most explosive period in AI history — is the one doing this? That's a signal that the talent exodus from the big labs isn't just about ego or money. It's about a different philosophy: build AI that people can own, customize, and control, rather than AI that people can only rent.

That's the kind of AI worth paying attention to.

---

If this kind of analysis is useful to you, CopperRiver runs on the same open-source models we write about — GLM, DeepSeek, Qwen, and now potentially Inkling. It's a desktop AI assistant that browses the web, runs terminal commands, reads your files, and automates real work. No cloud required for the sensitive stuff. Plans start at $9/month.