ThesisIs Inkling AI the Ultimate Open Source Model? Full Test teaches a practical creative automation move: A hands-on test of Inkling, a 975B-parameter (41B-active) mixture-of-experts model built for agentic coding and tool use, covering its self-fine-tuning demo, native multimodal reasoning, and three live tests: a Kanban app build, a calibrated Mars-2035 probability forecast, and a formatted long-form writing task.
The goal is not to remember the video. The goal is to extract the operating principle, tie it to timestamped evidence, test how far the claim transfers, and make something reusable.
0:00Sparse but Massive
“Writing code with AI is great, but getting it to use external tools is much harder. Today we are looking at Inkling, a newly dropped model that excels at agentic workflows. Let us look at the sheer scale...”
Inkling is a 975B-parameter mixture-of-experts model that only activates 41B parameters per query (with a lighter 'Inkling small' variant using 12B active parameters), supports a 1M-token context, and reasons natively over text, image, and audio without converting audio to text first; it was even used to fine-tune itself entirely on its own, writing its own synthetic training data to learn a constrained 'no letter E' language task. Write down Inkling's active-vs-total parameter split (41B of 975B) next to a model you already use, to build intuition for how mixture-of-experts routing changes the hardware you actually need.
4:42Agentic Coding Muscle
“functional web application in single short just from a simple text prompt. And then it used an embedded browser agent to actually click and interact with the final application. But it is not just good at single prompt...”
Inkling was purpose-built for agentic coding and tool use, with official demos showing it build a full web app from a single prompt and then use an embedded browser agent to click through and test it, plus a multiplayer snake game with real-time server bots; it matches Nemotron 3 Ultra's score on the Terminal Bench 2.1 coding benchmark while using only a third of the tokens, and it's deployable locally via SGLang, VLLM, or Unsloth, with a special NVFP4 checkpoint for Blackwell GPUs and a 98% strong-reject safety score. Pick one of the three deployment paths (SGLang, VLLM, Unsloth) and read its docs for loading a model with an adjustable thinking-effort setting, since that's the lever Inkling exposes for trading cost against quality.
7:40Calibrated Forecasting
“handle all the complex logic and the strict constraints that I had given in the initial prompt. And the second task was to check whether the human will land on Mars and return to Earth by 2035. And...”
Asked to forecast the probability of a human Mars round trip by 2035, Inkling didn't just guess: it gave roughly a 1% probability, walked through launch-window arithmetic, SpaceX's 'no earlier than 2028' Starship timeline, and life-support and certification requirements, then explicitly stated what evidence (like reusable orbital refueling by 2028) would raise its estimate above 10%. Give a reasoning model your own uncertain prediction question and require it to state both a probability and the specific evidence that would change that probability, then check whether its stated conditions are falsifiable.
01Brief
Start with this video's job: A hands-on test of Inkling, a 975B-parameter (41B-active) mixture-of-experts model built for agentic coding and tool use, covering its self-fine-tuning demo, native multimodal reasoning, and three live tests: a Kanban app build, a calibrated Mars-2035 probability forecast, and a formatted long-form writing task. Treat "Brief" as the outcome you are trying to make visible, not a topic label. Anchor it to 0:00, where the video says: “Writing code with AI is great, but getting it to use external tools is much harder. Today we are looking at Inkling, a newly dropped model that excels at agentic workflows. Let us look at the sheer scale...”
Source
Use "Source" to locate the part of the creative automation workflow the video is demonstrating. Ask what changes in your real setup if this claim is true. Anchor it to 4:42, where the video says: “functional web application in single short just from a simple text prompt. And then it used an embedded browser agent to actually click and interact with the final application. But it is not just good at single prompt...”
Generation
Turn "Generation" into the reusable artifact for this lesson: A creative workflow board with critique criteria and review checkpoints. This is where watching becomes something you can inspect and reuse.
Selection
Use "Selection" as the application surface. Decide whether the idea touches a browser flow, a local file, a model choice, a source document, a UI, or a review step.
Edit
Use "Edit" to prove the lesson. The evidence should connect back to the video title, transcript anchors, and a concrete output, not a generic best-practice claim.
Taste Review
Use "Taste Review" to carry the idea forward: save the prompt, checklist, diagram, or operating rule that would make the next agent run better.
ExampleSource-backed work packet
Convert the video into a scoped task that includes the transcript claim, target workflow, acceptance criteria, and proof. The output should be a creative workflow board with critique criteria and review checkpoints..
ExampleClaim vs. demo brief
Separate what the speaker claims, what the demo actually proves, and what still needs outside verification before you adopt the workflow.
ExampleTeach-back module
Transform the lesson into a definition, a mechanism diagram, one misconception, one practice exercise, and a check-for-understanding question.
Do not learn it wrong- Treating the title as the lesson without checking what the transcript actually says.
- Letting the prompt drift into generic advice that could apply to any video in the playlist.
- Copying the tool setup without identifying the operating principle that transfers to your own stack.
- Skipping the artifact, which means the learning never becomes operational or inspectable.
