By Thinking Machines LabLearning Objectives
- Understand what Inkling is and why Thinking Machines shipped open weights before a closed flagship
- Read the lab's own positioning critically — where Inkling competes and where it deliberately does not
- Identify when a customizable open base beats renting a stronger closed model
What Is Inkling?
Inkling is the first in-house model from Thinking Machines Lab, the research company founded by former OpenAI chief technology officer Mira Murati. Released on July 15, 2026, it is a multimodal mixture-of-experts (MoE) system with 975 billion total parameters that activates only about 41 billion for any given token. It was trained on 45 trillion tokens spanning text, images, audio, and video, and reasons across all four natively rather than bolting perception onto a text model. The context window reaches one million tokens, and the whole thing ships under an Apache 2.0 license with weights on Hugging Face.
The launch is unusual for what the lab refuses to claim. Most frontier releases arrive with a leaderboard chart and a superlative. Thinking Machines states plainly that Inkling is "not the strongest overall model available today, open or closed" — and positions it instead as a strong, open base that organizations fine-tune to their own data through Tinker, the lab's model-customization platform.
💡Key Concept
Mixture-of-experts (MoE): rather than running every parameter for every token, an MoE model routes each token to a small subset of specialized sub-networks called experts. Inkling holds 256 routed experts plus 2 shared ones and activates 6 per token — so it carries the knowledge of a 975-billion-parameter model while paying the compute cost of roughly a 41-billion-parameter one. This is what makes a model this large practical to self-host.
✅Tip
Access Inkling: full weights are on Hugging Face under Apache 2.0 for self-hosting. Fine-tuning runs on Tinker, and managed inference is available through Databricks and Baseten.
Key Capabilities
A Thinking-Effort Dial
Inkling lets callers turn reasoning effort up or down rather than exposing separate fast and slow models. Low effort trades depth for latency and cost; high effort spends more tokens on deliberation. The benchmark figures the lab published are measured near maximum effort, which is worth remembering when comparing them against models quoted at default settings.
Calibration Over Confidence
The lab tuned Inkling to give calibrated answers — flagging uncertainty rather than guessing fluently. That is a deliberate trade against benchmark scores, since a model that declines to guess forfeits the points it would occasionally win by luck. For applications where a confident wrong answer is worse than an admitted unknown — compliance, clinical triage, financial review — that trade is the point.
Native Multimodality
Text, image, audio, and video are all in the training mix, so Inkling handles them as first-class inputs. Its 91.4% on VoiceBench is the clearest evidence that the audio path is real rather than a wrapper around a separate speech model.
The Honest Benchmark Picture
Thinking Machines published these figures at near-maximum thinking effort:
| Benchmark | Inkling | What it measures |
|---|---|---|
| AIME 2026 (math competition) | 97.1% | Competition-level mathematical reasoning |
| GPQA Diamond | 87.2% | Graduate-level science questions |
| SWE-Bench Verified | 77.6% | Real-world software engineering tasks |
| MMMU Pro (multimodal reasoning) | 73.5% | College-level reasoning across images and text |
| VoiceBench | 91.4% | Spoken-language understanding |
| Humanity's Last Exam, with tools | 46.0% | Frontier-difficulty questions across domains |
These are strong numbers that would have led the field a year ago. They are also, as the lab concedes, not the best available today — GPT-5.6 and the current Claude flagships score higher on several of these. The relevant comparison is not against closed frontier models but against other open weights: Inkling is the most serious open-weights entry from a US lab, landing in a field that Chinese models like Kimi and Nvidia's Nemotron releases had largely defined.
⚠️Warning
Vendor-reported, at maximum effort. Every figure above comes from Thinking Machines' own launch materials, measured at an effort setting of 0.99. Treat them as claims until independent evaluations reproduce them, and expect lower scores at the effort levels most applications actually run at.
Inkling-Small
A preview variant, Inkling-Small, activates 12 billion of 276 billion total parameters and posts comparable scores on many benchmarks at lower latency and cost. For teams whose fine-tuning budget will not stretch to the full model, it is the more realistic starting point.
Why Open Weights First
The strategy only reads as modest if you miss the business model. If the product is the customization layer, the base model does not need to win outright — it needs to be strong, open, and adaptable enough that a fine-tuned Inkling beats a closed general-purpose model on the one narrow task a customer actually cares about. Giving the weights away seeds that pipeline; Tinker monetizes it.
The open question is demand. Fine-tuning takes data, evaluation discipline, and applied talent that most organizations do not have in house — which is the same gap that makes renting a closed frontier model the easy default.
Pricing
- Full weights on Hugging Face
- Self-host or fine-tune anywhere
- Commercial use permitted
- Fine-tune on Thinking Machines infrastructure
- Control over training algorithms and data
- 50% discount for a limited time at launch
- Available on Databricks and Baseten
- Managed inference without self-hosting
The Apache 2.0 license is the headline: no usage caps, no per-token bill, and commercial deployment permitted. The real cost of running Inkling is the compute to serve 41 billion active parameters plus the talent to fine-tune it well.
Inkling vs. Other Open-Weight Models
| Model | Lab | Open license | Positioning |
|---|---|---|---|
| Inkling | Thinking Machines | Apache 2.0 | Multimodal base built to be fine-tuned via Tinker |
| Kimi K2.6 | Moonshot AI | Open weights | Strong Chinese open-weights flagship |
| Llama 4 | Meta | Community license | The most widely deployed open-weights family |
| DeepSeek R1 | DeepSeek | MIT | Reasoning-focused open model |
| Gemma 4 | Google DeepMind | Open weights | Small-footprint open models from a frontier lab |
Related Tools
- Hugging Face Hub — where Inkling's weights are hosted
- Kimi K2.6 — the open-weights flagship Inkling is measured against
- Llama 4 — the incumbent open-weights family
- GPT-5.6 — the closed frontier model Inkling does not claim to beat
Strengths
- Apache 2.0 — no usage caps, no per-token bill, commercial deployment permitted, full weights downloadable
- Efficient for its size — 975 billion parameters of knowledge at roughly the compute cost of 41 billion active
- Genuinely multimodal — text, image, audio, and video in the training mix, with a 91.4% VoiceBench score backing the audio path
- Built to be customized — a thinking-effort dial, calibrated uncertainty, and a first-party fine-tuning platform in Tinker
- Honest positioning — the lab tells you where it does not lead, which is rarer than it should be
Limitations and Considerations
- Not the strongest model available — Thinking Machines says so directly; if you want maximum capability out of the box, a closed frontier model still wins
- Vendor-reported benchmarks at maximum effort — measured at effort 0.99 and not yet independently reproduced
- Self-hosting is not free — serving 41 billion active parameters needs real infrastructure, and the license does not pay for GPUs
- The value depends on fine-tuning — an un-customized Inkling is a weaker model than the closed alternatives it costs less than
- Young lab — founded in 2025, with a smaller support and tooling ecosystem than OpenAI, Anthropic, or Google
Key Takeaways
- Inkling (July 15, 2026) is Thinking Machines Lab's first model — a multimodal mixture-of-experts system with 975 billion total parameters, about 41 billion active per token, trained on 45 trillion tokens
- It ships under Apache 2.0 with weights on Hugging Face, a context window up to one million tokens, and a dial for reasoning effort
- The lab states plainly that Inkling is not the strongest model available, open or closed — it is pitched as a base to fine-tune through Tinker, not a leaderboard entry
- That candor is strategy, not modesty: if the product is customization, the base only needs to be strong enough that a fine-tuned version wins the narrow task a customer cares about
- The open question is whether enough organizations have the data and applied talent to fine-tune well, or default to renting a closed model instead
