Inkling 2 Small Weights: Licence, Download and Hardware Thinking Machines confirmed in a November 12 update to its launch post that the weights for Inkling 2 Small are now downloadable, moving the model beyond its earlier Tinker-only introduction. The release covers a model with 594 billion total parameters and 17 billion active parameters, and the official Hugging Face repository labels it apache-2.0, though the root file listing checked November 17 contained no standalone LICENSE file. The company's update lets teams evaluate a self-operated deployment, but the article notes that "Small" does not mean the model fits on a typical workstation, and that the main checkpoint and an MXFP8 alternative have different sizes and serving requirements. Inkling 2 Small’s weights are now downloadable. Thinking Machines added a November 12 update https://thinkingmachines.ai/news/inkling-2-small/ to its launch post confirming the release, following the earlier Tinker-only introduction. The practical change is that teams can evaluate a deployment they operate themselves. The word Small should not be read as a promise that the model fits on a typical workstation. The model has 594 billion total parameters and 17 billion active parameters. That active count describes the portion used in computation, not the complete storage requirement. A useful self-hosting decision needs three separate checks: which artifact you will download, whether the serving stack supports it and whether the available hardware can run the intended workload with enough memory left for requests. 1. 01The weights release is confirmedThe vendor’s November 12 update makes the download available rather than merely promising a future release. 2. 02Active parameters are not memoryThe full checkpoint and the runtime’s working memory matter when sizing a deployment. 3. 03Precision changes the artifactThe main checkpoint and the MXFP8 alternative have different sizes and serving requirements. 4. 04A hardware floor is not a complete recipeMatch the checkpoint, runtime build, GPU arrangement and expected request load before committing infrastructure. 01 — Release statusSeparate the new download from the earlier promise The earlier Inkling 2 Small article https://www.digitalapplied.com/blog/inkling-2-small-tinker-open-weights covered the model before its weights were released. The November 12 update changes that status. It does not mean a team should skip the practical review of the artifact now available. Hosted access, downloadable weights and a production-ready local service are different milestones. For a team that needs to experiment with a model under its own operational control, the release creates a concrete option. The first deliverable should be a reproducible evaluation using a specified checkpoint and serving stack. A download by itself proves neither that the model meets the task’s quality requirements nor that the organisation can operate it economically. Keep the original announcement date and the later weights-release date distinct in internal records. If a procurement note was written when weights were still promised, update its conclusion rather than leaving a future-tense assumption in place. At the same time, do not turn the new availability into a claim that every runtime or hardware configuration is immediately supported. The release removes the need to wait for the promised model download. It does not remove the need to verify the checkpoint, terms, serving software and useful performance on the intended workload. 02 — Licence evidenceRead the published label and preserve the artifact record The official Hugging Face repository https://huggingface.co/thinkingmachines/Inkling-2-Small labels the model apache-2.0, consistent with the licence promised in the launch post. The root file listing https://huggingface.co/thinkingmachines/Inkling-2-Small/tree/main checked November 17 did not contain a standalone LICENSE file. The supported statement is therefore that the publisher labels the release Apache 2.0; this article does not claim a separate repository licence file was inspected. Before redistributing an artifact or incorporating it into a product, preserve the exact repository revision, model card and applicable published terms with the deployment record. If the intended use depends on a detail that the available material does not resolve, obtain that clarification from the publisher. A concise record of the evidence is more useful than treating a licence badge as a complete procurement review. This also matters for derivatives. A community conversion, a quantised copy and the publisher’s own checkpoint are different artifacts even when they share a familiar model name. Check who produced the file and what accompanies it. Do not infer that every download carrying Inkling in its title has the same provenance, contents or deployment behavior. Keep the repository, revision, checkpoint format and accompanying licence information together. That makes later updates and redistribution decisions traceable without pretending that an unlocated file was reviewed. 03 — Model sizeDo not turn 17 billion active parameters into a RAM estimate Inkling 2 Small uses a mixture-of-experts design: only part of the model participates in processing a given token. The active-parameter count can help explain computation, but the larger collection of weights still has to be stored and made available to the serving system. It is therefore a poor shortcut for estimating the memory of a deployment. Consider a hypothetical infrastructure request based on the statement that the model is only 17B active. If the team orders hardware as though it were a conventional 17B-parameter checkpoint, the mistake appears before useful evaluation even begins. The right input to sizing is the actual artifact and a supported runtime recipe, followed by the memory needed for context, concurrency and execution. The local model deployment guide https://www.digitalapplied.com/blog/local-llm-deployment-privacy-guide-2025 provides broader context for controlling a model’s operating environment. For this release, keep the arithmetic tied to the published checkpoint rather than a simplified parameter label. A model can be efficient per generated token while still requiring substantial aggregate memory and a capable interconnect. How the model is described These describe architecture and computation. They are not a complete statement of the memory a serving process will consume. What is stored The downloaded artifact establishes the weight-storage burden. Different precision choices change that burden. What can be served The serving stack also needs space for context and execution. Concurrency can change the capacity requirement substantially. 04 — Checkpoint choiceDistinguish the main release from MXFP8 The SGLang recipe retrieved November 17, 2026 describes the main checkpoint as a mixture of numeric formats: routed experts in layers 0–39 use NVFP4, while layers 40–47 use MXFP8. It gives an approximate size of 394 GB. The separate Inkling-2-Small-MXFP8 checkpoint stores all routed experts in MXFP8 and is about 623 GB. Those names should be preserved in the evaluation record. A comparison between two deployments that silently uses different checkpoints is not a clean comparison of hardware or runtime settings. The smaller artifact may be the practical starting point, but the decision should follow the publisher’s guidance and the requirements of the task, not an assumption that a higher-precision label always produces a meaningful business improvement. The main repository’s file listing includes 24 model shards and a separate MTP file of about 3.42 GB. Download planning should account for the actual files required by the selected recipe. Avoid treating an approximate repository total as a precise RAM requirement: disk storage and the runtime’s loaded representation are related, but they are not the same measurement. | SGLang checkpoint descriptions and official repository listing, retrieved November 17, 2026. Sizes describe artifacts, not a measured peak-memory or throughput result from Digital Applied. | | | |---|---|---| | Checkpoint | Published approximate size | Key distinction | |---|---|---| | Inkling-2-Small | 394 GB | Mixed NVFP4/MXFP8 routed experts | | Inkling-2-Small-MXFP8 | 623 GB | All routed experts in MXFP8 | 05 — GPU requirementsStart with the vendor floor, then check the recipe The vendor’s model card specifies at least 500 GB aggregate VRAM for the main checkpoint, with examples of two B300 or four H200 GPUs. For MXFP8, it specifies at least 800 GB, with examples of four B300 or eight H200 GPUs. These are publisher requirements and example configurations, not a Digital Applied hardware test. A memory total is still not a complete serving design. GPU type, parallelism, runtime support, context length and request concurrency affect what a configuration can do. A vendor example showing enough aggregate memory does not establish that every framework recipe uses that exact GPU count or that the system can sustain the traffic your application expects. Before reserving infrastructure, choose a documented recipe for the intended checkpoint and hardware family. Confirm the supported runtime build and expected operating mode, then run a bounded acceptance test. If the published hardware floor and a framework’s validated example differ, do not silently substitute one for the other. Explain which source supports the configuration you actually intend to use. | Thinking Machines model-card guidance, retrieved November 17, 2026. These are vendor-stated floors and examples; runtime-specific recipes and workload capacity must be checked separately. | | | |---|---|---| | Checkpoint | Vendor minimum aggregate VRAM | Vendor example configurations | |---|---|---| | Main checkpoint | 500 GB | 2× B300 or 4× H200 | | MXFP8 alternative | 800 GB | 4× B300 or 8× H200 | 06 — Serving softwarePin the supported build instead of assuming latest works The SGLang recipe checked November 17 says Inkling support is not yet in an SGLang release and directs users to development images. That is an important qualification for an infrastructure plan. A generic installation command or a familiar latest tag may not reproduce the model-specific support described by the recipe. Create a small deployment manifest that records the checkpoint revision, container or package build, hardware arrangement and relevant parser settings. Keep it with the evaluation results. If a test later fails after an update, that record helps distinguish a changed model artifact from a changed serving stack. Without it, the team may spend time comparing outputs from environments that were never equivalent. A development build also changes the operating expectation. Use an isolated trial and a rollback route before making it a shared service. Test startup, a representative request, a failed request and a restart. The purpose is not to prove every feature in the framework; it is to establish that the selected combination can reliably perform the task you plan to support. Pin the model revision and runtime build, then retain the exact configuration with the result. “It worked with the newest image” is difficult to investigate once that image changes. 07 — Quality checksEvaluate the modality and behavior the application needs The model accepts text and image input and generates text. That is useful for tasks such as interpreting a screenshot alongside a written requirement, but it does not make the model an image-generation service. Define the expected output before designing the evaluation, and include cases where the image is unclear or the necessary information is absent. For a hypothetical document assistant, a useful test might require the model to distinguish a visible figure from a conclusion that cannot be drawn from the page. For a coding agent, it might require the agent to inspect the actual error and return a minimal change with evidence. These examples test application behavior rather than assuming that broad model capability guarantees a suitable workflow. The agent evaluation guide https://www.digitalapplied.com/blog/ai-agent-eval-frameworks-testing-guide-2026 helps separate answer quality from complete-run reliability. Run the same acceptance set against the exact checkpoint and runtime that would be deployed. If a serving change alters tool parsing or output formatting, it can matter operationally even when a short conversational response still looks fluent. Include a task whose correct result is “the source does not establish this.” An honest limit is often more useful than a confident answer assembled from missing evidence. 08 — Deployment choiceChoose self-hosting for a reason the team can sustain Downloadable weights make self-hosting possible, but possibility is not the same as an economic recommendation. Compare the expected workload with the cost of keeping the selected infrastructure available, monitoring it and updating the serving stack. Include idle capacity and operational effort. A low marginal token cost does not automatically make a lightly used cluster economical. Fine-tuning adds another decision rather than replacing the inference review. The model’s open-weight status supports downstream experimentation, but a successful inference configuration does not establish the memory or software needed for a training job. Define the desired improvement, the evaluation data and the rollback plan before committing to a custom model variant. AI transformation services https://www.digitalapplied.com/services/ai-transformation should connect the hosting choice to a concrete business requirement: control over deployment, a repeatable specialised task or another benefit the team can verify. Begin with the supported artifact and one useful evaluation. Expand only when the quality and operating evidence justify the larger commitment. Approve a bounded evaluation with a specified checkpoint and serving recipe. Approve ongoing infrastructure only after the team can explain the useful capacity, quality and operating cost it provides. Verify the artifact before reserving the cluster The weights release gives Inkling 2 Small users a real self-hosting choice. Make that choice from the checkpoint, published licence information, supported runtime and workload capacity, rather than the model’s name or active-parameter count. A reproducible evaluation is the bridge between an available download and a service worth operating.