AI News, 4 October: Whistle, FLUX 3 Image and Comfy Agent

01Early-October tools, caught up today

Cactus released Whistle on October 2, making short speech transcription the freshest tool in this issue. FLUX 3 Image and Comfy Agent were both announced on October 1 and are included as useful catch-up, not October 4 launches. Together they address three distinct production bottlenecks: capturing a spoken instruction on a device, placing or revising visual elements, and building the node graph that runs a creative workflow.

02Whistle: short speech inputs without a server round trip

Whistle is Cactus's 16.9 MB speech-recognition model for its Needle CPU engine. It accepts 16 kHz mono audio and transcribes English, German, French, Spanish, Italian, Dutch and Polish. It can return word timestamps for highlighting or seeking, and speech embeddings, numerical audio features, without decoding text. The same engine can load Whistle beside a Needle language model, so a device can turn a short spoken request into a tool call without handing a transcript to a separate service.

A wearable or field recorder could capture an instruction, show the recognized words and let the user correct them before any action runs. That is an editorial use case, not an XANTRANS test. Cactus's latency and recognition comparisons are vendor measurements on specified hardware and datasets, not a guarantee for a noisy room or different processor.

The model handles up to 30 seconds in one pass and caps a transcript at 320 tokens. Latvian is not among its seven supported languages. Cactus provides Apache 2.0 model weights and a Python starting point with pip install cactus-needle; the base install takes 16 kHz WAV or raw samples, while microphone capture and other sample rates need its extra dependencies. No hosted-service price is established here.

03FLUX 3 Image: specify the layout, then target the edit

Black Forest Labs released FLUX 3 Image on October 1. Its image API uses one endpoint for generation and editing, accepts up to ten reference images, and lets a prompt end with bounding boxes that specify where elements belong. The boxes use a 0–1000 coordinate grid. They can also target a local change such as moving, replacing or recoloring one element while the surrounding composition is intended to stay put. Output options reach the 4k tier, roughly 16 megapixels.

For an editorial cover, a designer could position a headline, face and object as separate regions, then request a revision to the object without starting the whole layout again. For an ad, references could carry a product and a scene into one composition. These are possible workflows, not examples XANTRANS generated. Local-edit fidelity and typography still need visual comparison against the approved source; a box describes intent rather than proving that every untouched pixel survived.

You can try the browser Playground or use POST /v1/flux-3-image with a BFL API key. BFL's published per-image API rates by resolution are $0.041 for 768sq, $0.048 for 1k, $0.100 for 2k and $0.607 for 4k, checked October 4. These are API resolution rates, not a verified Playground allowance or self-hosting price. BFL directs commercial-weight inquiries to sales.

04Comfy Agent: edit the graph while the artist stays on the canvas

Comfy introduced Comfy Agent on October 1. Inside ComfyUI, a request can ask it to find a suitable template or nodes, assemble a graph, change a specific part, validate connections and run the workflow. It can inspect image contents and video or audio properties, use reusable skills and keep several chat sessions open. The artist can still inspect and change the nodes while the agent works, which matters when a generation needs careful model and parameter choices.

A studio could ask for three versions of an existing video workflow with different models, then compare the graphs and outputs before selecting one. The request should name what must remain fixed, such as source media, aspect ratio or an approved prompt. Comfy's examples show what the agent is designed to do; XANTRANS has not run that comparison. Generated nodes and outputs still need review, particularly where missing inputs, paid models or changed settings affect the result.

Comfy Agent is available now in Comfy Cloud. Support inside local Comfy Desktop is described as coming soon, and agent requests use Comfy's network services. Initial chats or tokens are free to try; after that, usage draws from Comfy Credits, with token-based rates in its documentation. Running the resulting generation can consume credits too. It is neither an unlimited free assistant nor an available fully local workflow today.

05What to try first

Choose Whistle when short spoken input and on-device processing matter, but check the language and clip length first. Choose FLUX 3 Image when layout placement or a focused revision is the main job, and budget for the output resolution you need. Choose Comfy Agent when the bottleneck is wiring or revising a ComfyUI graph, with Cloud access and credit use understood before running generations.