Inkling has open weights, but 41 billion active parameters do not fit on a laptop

On July 15, 2026, Thinking Machines Lab released Inkling with downloadable weights under the Apache 2.0 license. The model accepts text, images and audio, generates text and supports a context window of up to one million tokens. Its headline number—975 billion parameters—comes with another figure that better explains the architecture: about 41 billion are active for each token.

Calling it simply “open” hides several decisions. The company supplies weights and permits modification, but it does not publish the 45-trillion-token training set or make the model inexpensive to run. The original Inkling announcement accurately calls it open-weights and acknowledges that it is not the strongest model available, open or closed. Its proposition is a customizable multimodal base. The durable skill is translating “open weights” into specific rights, artifacts, resources and responsibilities.

975 billion parameters do not work at once

Inkling is a mixture-of-experts, or MoE, transformer. Each sparse layer contains 256 routed experts and two shared experts. For every token, a router selects six of the 256 while the two shared experts remain active. The model can therefore store enormous total capacity without multiplying through every parameter at every step.

The 41-billion active figure helps compare per-token computation, but it still does not summarize total cost. Attention layers, multimodal representations, context cache, accelerator communication and system overhead remain. All weights must also be available in memory or distributed across machines even when only a fraction is selected for a token.

Routing introduces a further systems question: selected experts may sit on different accelerators. Keeping experts balanced and moving intermediate states across a fast interconnect can affect throughput as much as the advertised arithmetic. A procurement comparison should therefore request measured requests per second, time to first token, output speed, batch assumptions and power—not infer serving economics from active parameters alone.

The official model card gives the practical consequence: the BF16 checkpoint requires at least 2 TB of aggregate VRAM, with listed configurations of eight B300 GPUs or sixteen H200s. The NVFP4 version lowers the requirement to 600 GB, still using four B300s or eight H200s depending on mode. “Downloadable” does not mean “runs on a laptop.”

Quantization reduces the numerical precision used to store and operate on weights. It may save memory and accelerate inference, but quality equivalence must be tested on the user’s task. Forty-one billion active describes sparsity; 2 TB and 600 GB describe deployment. Combining those measures creates the false impression that the system occupies the same space as a dense 41-billion-parameter model.

Open weights do not mean an open process

There are at least five layers of openness. First, access: can the weights be downloaded? Second, permission: does the license allow use, modification and redistribution? Third, reproducibility: are data, code and recipe sufficient to repeat training? Fourth, transparency: are evaluations, limits and provenance described? Fifth, operability: are the required hardware and software within reach?

Inkling is strong on the first two. The weights are in the official Hugging Face repository, and the listed license is Apache 2.0. The card says training data came from publicly available sources, third-party acquisitions and synthetic or augmented material, with cleaning and filtering. It does not identify the complete corpus or provide an end-to-end reproducible recipe.

That does not invalidate the release. It does prevent “open source” from serving as shorthand for total transparency. An organization can retain, adapt and run the weights without depending on the creator’s inference API, while still needing to examine rights in its own fine-tuning data, software dependencies, hardware cost and behavior inherited from pretraining.

Multimodal input, textual output

The model processes UTF-8 text, images and 16 kHz WAV audio; the declared output is text. The card describes a hierarchical image patch encoder and discrete audio encoding, both projected into the decoder’s shared space. Video appears among training-data types, not as a declared input modality in the model card.

This distinction avoids another common shortcut. “Trained on video” does not mean “accepts video,” and “multimodal” does not mean the model produces images, speech or video. Before designing an application, write an input-output matrix covering format, limits, resolution, duration, language and result. With Inkling, an image or recording can inform reasoning, but the base model’s product remains text.

A maximum context window must also be separated from effective memory. One million tokens states an accepted limit, not that every detail at the beginning is recalled with equal accuracy or that filling the window is economical. The relevant test uses real documents, varied positions and distractors while measuring cost and latency.

How to read its results table

The card compares Inkling with open- and closed-weight models across reasoning, code, agents, factuality, chat, vision, audio and safety. It reports, for example, 77.6% on SWE-bench Verified, 87.2% on GPQA Diamond and 43.9% on SimpleQA Verified. The spread matters more than selecting the highest number: one system can be competitive on one task and far behind on another.

The principal column uses effort=0.99, a high reasoning-effort setting. Cost comparisons require aligned effort, output length, infrastructure, batch size and quantization. A claim of token efficiency may hold in one setup, but “fewer tokens” alone does not prove lower latency or total cost. More expensive tokens, a larger cache or cross-node communication can reverse the result.

The evaluations are published by the provider. Notes disclose some internal harnesses and special treatments, which is better than leaving them hidden, but buyers still need reproduction. An enterprise test should contain representative cases, a held-out set not used for tuning, blind judging where feasible, and measurements of quality, cost, latency and failure modes.

Customization transfers control and responsibility

Thinking Machines makes Inkling available for fine-tuning through Tinker. Customization can teach format, vocabulary, tool use or a response policy. It does not automatically inject all of an organization’s truth, and training on documents is not a replacement for retrieval when information changes. Nor does it remove hallucinations: it can reinforce errors in the example set.

A sound pilot starts by defining which behavior should change. Training, validation and test sets remain separate, with independent documents or customers where appropriate. Then compare the base model with the adapted one and include regression tests for safety, languages, legitimate refusals, stale information and out-of-distribution inputs.

Downloading also changes governance. Local operation can reduce exposure to an external provider, but the operator takes responsibility for patches, access controls, logging, extraction resistance, filters and incident response. If an adaptation is redistributed, license notices must be preserved and changes documented. API independence does not mean freedom from obligations.

The decision card that fits on one page

For any open-weight model, record the exact license; downloadable artifacts; total and active parameters; memory by format; input and output modalities; tested context; disclosed data provenance; reproducible evaluation; cost on your infrastructure; and the owner of safety after adaptation. A field without a source remains unknown.

Inkling is a substantial release: full weights, a permissive license, documented architecture and compatibility with several inference engines. It is also a system that needs cluster-scale infrastructure at full precision and whose corpus cannot be publicly reproduced. Holding both statements together is more useful than arguing whether it is “open” or “closed.” Openness is not a binary label; it is an auditable stack of permissions and dependencies.