Inkling: Thinking Machines' Open-Weights Gambit

The model

Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active per token. It was pretrained on 45 trillion tokens across text, images, audio, and video, and supports a 1M-token context window. Full weights are on Hugging Face under Apache 2.0. A smaller variant, Inkling-Small (276B total, 12B active), is previewed with weights coming soon.

The benchmark story is mixed. SWE-Bench Verified sits at 77.6%, AIME 2026 at 97.1%. These are solid numbers for a first release. But on HLE text-only (29.7%), it trails Kimi K2.6 (35.9%), GLM 5.2 (40.1%), and every closed frontier model. SimpleQA Verified at 43.9% is unremarkable. The blog post itself is unusually honest: "Inkling is not the strongest overall model available today, open or closed."

The "outside of China" problem

Deedy's framing, "best open weight AI model outside of China," immediately drew fire. One reply: "It's like saying best if you exclude glm, deepseek, kimi, qwen and a bunch of others." Another: "Below Kimi 2.6 = 'on top'. Make it make sense."

The qualifier reveals something real. Chinese labs have been dominating open-weights releases for months. DeepSeek, Kimi, GLM, Qwen are all shipping competitive or superior models. When your headline achievement requires a geographic asterisk, you're acknowledging the gap rather than closing it. Deedy's follow-up is more grounded: "#1 outside of China is a pretty decent start but yes I imagine they're going to have to ramp compute pretty aggressively to get to a Kimi K3 / Kivine."

That said, the "outside of China" framing matters for one practical reason: many U.S. enterprises won't deploy Chinese models regardless of performance. Regulatory anxiety, data sovereignty, procurement policies. Inkling doesn't need to beat Kimi K2.6 on every benchmark to win enterprise deals. It needs to be the best model that a Fortune 500 CISO will actually sign off on. Right now, it might be.

The actual differentiator: controllable thinking effort

The most interesting thing about Inkling isn't any single benchmark score. It's the controllable thinking effort, which lets developers trade cost and latency against performance on a continuous scale. Thinking Machines reports that Inkling spends one third as many tokens as Nemotron 3 Ultra to reach the same Terminal Bench 2.1 score. On the Artificial Analysis token efficiency metric, Inkling uses roughly 25K output tokens per task versus GLM 5.2 at 43K, Kimi K2.6 at 38K, and DeepSeek V4 Pro at 37K.

This matters more than peak scores for the use case Thinking Machines is targeting. If you're running a model millions of times inside an agent loop, or paying for inference at scale, token efficiency compounds. A model that gets 90% of the way there at one third the cost is more valuable than one that gets 95% at full cost.

The RL training revealed something fascinating here. Over 30M rollouts, the chain of thought became more concise on its own, dropping grammatical overhead while remaining comprehensible. Efficiency alone drove the compression. Nobody targeted it with a reward. The model learned to think faster because thinking faster was rewarded by the environment.

Audio is the sleeper capability

Inkling's audio numbers are genuinely strong. AudioMC at 56.6% beats Qwen3-Omni (24.3%) and Nemotron-3 Nano-Omni (23.2%) by enormous margins. MMAU at 77.2% is competitive with Qwen3-Omni (77.5%). VoiceBench at 91.4% beats both. Only Gemini 3.1 Pro (66.8% AudioMC, 82.5% MMAU, 94.3% VoiceBench) is clearly ahead.

Most open-weights models treat audio as an afterthought. Inkling treats it as a first-class modality, trained from scratch with an encoder-free architecture using dMel spectrograms. This connects directly to Thinking Machines' interaction models vision: real-time collaboration using voice and vision. If the future of AI interfaces is conversational rather than text-in-text-out, Inkling is one of the few open models built for it.

The business thesis: sell them their own models

Xiaoyin Qu's post cuts to the real strategic play. Her argument: Thinking Machines won't beat Anthropic by building a better frontier model. They'll beat them by becoming the Palantir FDE (forward deployed engineer) for enterprise custom models. The playbook is three steps: release the best American open-weight model, drive enterprise adoption, then charge seven to nine figures to post-train and run custom models behind enterprise firewalls.

The fat margin will move to customization. Meanwhile, Henry-ford-styled, standardized models will make no margins. OpenAI and Anthropic will have their API margins squeezed by Deepseek/GLM/Grok/Meta etc, and their consumer subscriptions are loss centers. @quxiaoyin on X

The thesis rests on enterprises wanting their own models, trained on their own data, continuously updated for different workflows. That creates sticky recurring revenue. This is where the open-weights decision makes business sense. You can't fine-tune a model behind your own firewall if the weights are closed. Anthropic and OpenAI's closed-weights strategy locks them into the API margin game. Thinking Machines' open weights make on-premise customization possible, and Tinker is the interface that makes it accessible. The model release isn't the product. The model is the wedge. Tinker is the product.

The pushback in Qu's replies is worth noting. Eric Seufter points out the overlap problem: companies that don't care about frontier performance AND can afford custom fine-tuning is a narrow segment. Someone asks why Palantir doesn't just win this themselves. And the open-source community may figure out fine-tuning without paying Thinking Machines for the expertise. The thesis is compelling but unproven. How many enterprises actually want to own models versus just calling an API?

Where this leaves the open-weights race

The Western open-weights landscape has been thin. NVIDIA's Nemotron models are competitive but not exciting. Google's Gemma line is smaller-scale. Meta's Llama releases have been quiet. gpt-oss-120b sits at 24 on the Intelligence Index. Into this gap, Thinking Machines drops a 975B multimodal model with genuine strengths in agentic coding, token efficiency, and audio.

Inkling doesn't need to be the best model in the world. It needs to be the best base model for customization that Western enterprises feel safe deploying. That's a narrower claim, but it's one Thinking Machines can defend. The combination of open weights, Tinker for fine-tuning, broad ecosystem support (Together, Fireworks, Databricks, vLLM, SGLang, Unsloth), and a multimodal foundation is a coherent package. Few others are offering all of it.

The risk is that Kimi K3 or DeepSeek's next release widens the gap again before Thinking Machines ramps compute. Deedy flagged this explicitly. Six months from now, "best outside of China" could sound even more like a consolation prize. But the business model doesn't require winning the benchmark race. It requires being good enough to fine-tune, efficient enough to run at scale, and open enough to deploy anywhere. Inkling checks those boxes on day one.