⏱️ Lectura: 8 min
On July 15, 2026, Thinking Machines Lab released the full weights of Inkling, a Mixture-of-Experts model with 975 billion total parameters and 41 billion active per inference.
📑 En este artículo
The model supports up to 1 million tokens of context and ships alongside Inkling-Small, a lighter version built to run at lower cost. Both can already be fine-tuned on Tinker, the company’s platform.
TL;DR
- Thinking Machines Lab launched Inkling on July 15, 2026, its first fully open-weights model.
- Inkling is a Mixture-of-Experts model with 975 billion total parameters and 41 billion active.
- It supports up to 1 million tokens of context and was pretrained on 45 trillion tokens of text, image, audio, and video.
- The company also introduced a preview of Inkling-Small, a version with 12 billion active parameters.
- Both models can already be fine-tuned on Tinker, the company’s platform.
- Thinking Machines demonstrated that Inkling can retrain itself: it wrote its own dataset and evaluated the result.
- The self-tuning process, turning Inkling into a lipogram model without the letter e, took about 27 minutes.
Introduction
Thinking Machines Lab defines its mission as building AI that extends human will and judgment. Until now that mission had shown up in three moves: a platform for customizing models (Tinker), a preview of an AI system designed for interactive collaboration, and a series of research publications. Inkling is the fourth move and the first to deliver a model trained from scratch with the full weights available for download.
The difference from a closed release is concrete: anyone can download Inkling’s weights, run them on their own infrastructure, and modify them, something that isn’t possible with a model accessed only through an API.
What Happened
Inkling is a Mixture-of-Experts (MoE) transformer with 975 billion total parameters, of which only 41 billion activate at each inference step. That ratio between total and active parameters is the key to the MoE architecture: instead of using the whole network for each token, a router selects a subset of experts, which cuts the compute per token without giving up the model’s total capacity.
The model supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens combining text, images, audio, and video. At inference time, Inkling reasons natively over text, image, and audio, and lets you control the thinking effort to balance cost against response quality.
Alongside Inkling, Thinking Machines Lab shared a preview of Inkling-Small, a variant with 12 billion active parameters, trained with a similar recipe but designed to run at lower cost and latency.
Context and History
The company is explicit about the goal of this release: Inkling doesn’t aim to be the strongest model on the market, open or closed. It aims to be a balanced, flexible base, useful for customization, with multimodal capabilities, efficient reasoning, and immediate availability for fine-tuning on Tinker.
That approach sets it apart from releases that compete solely for benchmark rankings. Thinking Machines is betting that combining open weights with an accessible fine-tuning platform (Tinker) will spawn an ecosystem of specialized variants rather than a single dominant generalist model.
Technical Details and Performance
The Mixture-of-Experts architecture splits the network into multiple subnetworks (“experts”) and a routing mechanism decides, token by token, which experts take part. With 975B total parameters and only 41B active, Inkling pays the compute cost of a model much smaller than its nominal size while keeping the ability to store knowledge distributed across all the experts.
The following table summarizes the differences between the two announced variants:
| Model | Active parameters | Context | When to use it |
|---|---|---|---|
| Inkling | 41B (975B total) | up to 1M tokens | best quality, complex multimodal tasks |
| Inkling-Small (preview) | 12B | same training recipe | lower cost and latency, rapid prototyping |
For local inference with the standard Hugging Face library, loading the model follows the usual transformers pattern:
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("thinkingmachines/Inkling")
model = AutoModelForCausalLM.from_pretrained("thinkingmachines/Inkling")
inputs = tokenizer("Explica que es un modelo Mixture-of-Experts", return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(output[0], skip_special_tokens=True))This snippet loads the tokenizer and the model, generates a response of up to 200 new tokens, and prints it as plain text. To confirm which variant ended up loaded in memory, just inspect model.config.num_parameters or the checkpoint name in model.name_or_path.
How to Get Started
There are two paths to try Inkling today. The first is to download the weights from Hugging Face and run local inference with transformers, as in the example above. The second is to go straight to the Inkling Playground, the chat interface Thinking Machines Lab added to the Tinker console so you can talk to the model before deciding whether it’s a good base for fine-tuning.
💡 Tip: before launching a long fine-tuning run, try the base model in the Inkling Playground: the feel of real usage complements what the benchmarks show.
Thinking Machines Lab showed the full fine-tuning flow using Inkling itself as the agent: the model wrote its training objective (becoming a lipogram model that never uses the letter “e”), defined a binary scoring function, launched the job on Tinker, and waited for the result. The evaluation function snippet it wrote, adapted from the original, is this:
OBJECTIVE = "Un modelo lipograma que evita la letra e en todas las respuestas."
GEN_SYS = (
"Sos un asistente lipograma. Debes escribir todas las respuestas sin usar la letra 'e' o 'E'. "
"Evita ese simbolo por completo. Escribi respuestas largas y naturales usando solo palabras sin 'e'."
)
THRESHOLD = 10.0
def score(prompt, answer):
if "e" in answer or "E" in answer:
return 0.0
return 10.0The scoring function is binary on purpose: if the response contains a single letter “e”, the job discards it with a score of zero. The full pipeline (train, evaluate against the base model, and update the weights) finished in about 27 minutes, according to the company’s own announcement.
The following diagram summarizes that self-tuning cycle:
flowchart TD
A["Inkling base"] --> B["Escribe su propio training job"]
B --> C["Ejecuta el fine-tuning en Tinker"]
C --> D["Evalua el resultado contra el modelo base"]
D --> E{"Supera el umbral fijado?"}
E -->|"Si"| F["Publica el checkpoint y actualiza sus pesos"]
E -->|"No"| BImpact and Analysis
For development teams in Latin America, an open-weights model of this size opens the door to running inference and fine-tuning without depending on a closed API, something relevant in contexts where billing in dollars or latency to distant servers is a real problem. The trade-off is that 975B total parameters, even if only 41B are active, still demand considerable memory infrastructure for serious self-hosting; that’s where Inkling-Small, with 12B active, becomes the practical choice for teams without access to large clusters.
💭 Key point: Thinking Machines Lab openly acknowledges that Inkling is not the strongest model available, open or closed. It’s betting on being a flexible base for customization, not on topping benchmark tables.
That editorial honesty from the company itself is a data point in its own right: in a market where every release tends to be sold as the best in its class, Thinking Machines Lab chooses to position Inkling by its adaptability (multimodality, controllable thinking effort, availability on Tinker) rather than by performance metrics.
What’s Next
The company confirmed that Inkling is the first in a family of models of different sizes. Inkling-Small is still in preview, which suggests adjustments before a stable release. It remains to be seen whether Thinking Machines Lab will publish intermediate variants between 12B and 41B active, and whether the Inkling Playground will add more evaluation tools inside the Tinker console.
📖 Summary on Telegram: View summary
Try it yourself: download Inkling’s weights from Hugging Face or open the Inkling Playground in the Tinker console and compare it against the model you use today.
Frequently Asked Questions
What is Inkling?
It’s the first fully open-weights model trained from scratch by Thinking Machines Lab, with a Mixture-of-Experts architecture.
How many parameters does Inkling have?
975 billion total parameters, of which 41 billion activate at each inference.
What is Inkling-Small?
A preview version with 12 billion active parameters, trained with a similar recipe but optimized for lower cost and latency.
Where can you try or fine-tune Inkling?
The weights are available on Hugging Face for local inference, and fine-tuning is done on the Tinker platform, which also includes the Inkling Playground for chatting with the model.
What does it mean that Inkling retrained itself?
Thinking Machines Lab showed a demo where Inkling itself wrote its fine-tuning objective, launched the job on Tinker, evaluated the result, and updated its weights, all in an agent session of about 27 minutes.
Is Inkling the most powerful AI model on the market?
No. Thinking Machines Lab itself clarifies that it’s not the strongest model available, open or closed: its value lies in being a flexible base for customization.
References
- Thinking Machines Lab: Introducing Inkling: official announcement with the model’s spec sheet and the self-tuning demo.
- Hugging Face: platform where Thinking Machines Lab published Inkling’s weights for download.
- Wikipedia: Mixture of experts: general explanation of the architecture Inkling uses.
- Thinking Machines Lab: the company’s official site, including the Tinker platform for fine-tuning.
📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day. @programacion
Imagen destacada: Foto de Steve A Johnson en Unsplash
0 Comments