Inkling powers podcast clipping app with FFmpeg edits

• The build turns Inkling into an audio-first clipping agent: venturetwins' demo has it listen to full podcast episodes, choose moments or search topics, then direct FFmpeg cuts.
• The pipeline is concrete: venturetwins said the app downloads a YouTube video, converts it to WAV, batches the audio, asks Inkling for candidates, checks boundaries, and renders with FFmpeg in their build notes.
• Native audio is the workflow hinge: venturetwins said Inkling can use vocal emphasis, emotion, and cadence in their follow-up on audio cues.
• Inkling fits the job because it accepts audio as a first-class input: Thinking Machines describes it as a 975B total parameter, 41B active MoE trained on text, images, audio, and video, while aakashgupta's thread framed the open model as a base for customization.
• Current scope is narrow: venturetwins said the app is “just audio and image rn” in one scope reply, and broader orchestration was not built in this version, according to another follow-up.

Thinking Machines’ Inkling announcement gives the raw shape: open weights, up to a 1M token context window, controllable thinking effort, and native reasoning over text, images, and audio. The audio docs are the builder-facing bit: WAV, MP3, and FLAC input, client-side DMel encoding, and no remote URL fetching. Hugging Face’s launch post adds the deployment context, with day-0 support in transformers, SGLang, and llama.cpp.

venturetwins' podcast clipping app demo

The app has Inkling listen to whole episodes and tell FFmpeg where to cut. The demo shows timestamped clip candidates, then a topic search that jumps to a relevant segment.

Podcast clipping is an especially good target for this kind of agentic media workflow. In a reply, venturetwins called clipping “the bane of basically every podcaster’s existence” in a podcaster reply.

venturetwins said the full app was built in Cursor. The loop is short enough to steal:

1. Download the YouTube video and convert it to WAV.
2. Split the long audio into batches.
3. Have Inkling reason across the audio and propose clip candidates.
4. Run a correction pass to check that each clip starts and ends in the right place.
5. Render the final clips with FFmpeg.

The correction pass is the quiet useful piece: Inkling chooses candidates, then a second model pass checks timing before deterministic rendering.

venturetwins said most podcast clipping tools they had tested convert audio to transcript first. Inkling’s advantage in this build is that it can use non-text signals: vocal emphasis, emotion, and speech cadence.

In another reply, venturetwins said they wanted a model that could “listen” instead of a transcript-based system in a transcript reply. Thinking Machines’ audio docs list the same kind of inputs as supported use cases:

• Speech transcription.
• Audio question answering.
• Speaking-style classification.
• Information extraction from recordings.

The docs also say remote HTTP and cloud-storage media are not fetched by the renderer. That matches the app’s first step: download the YouTube file locally, then feed Inkling a WAV.

The official announcement says Inkling is a sparse Mixture-of-Experts transformer with 975B total parameters and 41B active parameters. It was pretrained on 45 trillion tokens across text, images, audio, and video.

The same announcement says Inkling is not the strongest overall model available. That line matters because this app is not asking the model to win a chat benchmark, it is asking the model to inspect long raw media and hand structured decisions to a tool.

A few implementation details explain the fit:

• The model card says Inkling accepts text, image, and audio inputs and generates text outputs.
• The product page lists a 1M token context window, with 64K and 256K context options on Tinker.
• The tml-renderers docs say chat messages, tool calls, images, and audio are converted into the token and media representation Inkling expects.
• The audio docs say audio is decoded, resampled, and encoded as quantized mel features on the client.

LLMJunky called Inkling “one of the best US open models,” then clarified in replies that this did not mean it was the best US model in a wording reply. Another reply explained the table coloring: red marks the lowest result, green marks the highest, and the delta compares Inkling with the best result in each row in the table-key reply.

That framing lines up with Thinking Machines’ own positioning. Inkling is a competitive open base with unusual multimodal inputs, not a closed frontier model trying to dominate every column.

The current app is scoped to audio and image. venturetwins told one user it is “just audio and image rn,” and another reply said broader orchestration sounded possible but was not implemented in this version orchestration reply.

Inkling itself is also input-multimodal, not output-multimodal: the model card says it accepts text, image, and audio inputs and generates text outputs. A separate venturetwins reply was just a brief agreement with a commenter, not an implementation note.