# AI News, 17 October: Pika in ChatGPT and new open-weight models

Pika announced its ChatGPT creative integration on 16 October. This issue also catches up on four 14–15 October releases: FLUX 3 Image Klein, Muse Spark Open, Nemotron 3.5 and Ultramarine 1 Flash.

Published: 2026-10-17. Updated: 2026-10-17. Author: XANTRANS editorial.

Pika connects a creative brief to media tools inside ChatGPT. The model releases add options for image editing, visual agents and document work, with different credit, hardware and availability requirements.

## Pika brings media creation into ChatGPT

Pika announced its ChatGPT plugin on 16 October. It brings image, video, music, sound effects and speech tools into a conversation, alongside edits such as extending clips or changing an image. Tool choices and results appear inside the chat. A timeline lets users adjust timing, regenerate clips and add transitions or music, working with assets from their Pika library.

For a short product film, a team could start with a packshot and brief, create a draft clip, then request a different ending and soundtrack in the same workspace. Outputs still need review for visual continuity and suitability. Installation requires Pika sign-in and authorization. Installing does not itself generate media: Pika says the plugin normally shows an estimated cost before generation spends Pika credits. Available actions depend on account permissions and spending preferences. A ChatGPT conversation does not make the media generation free.

## FLUX 3 Image Klein: downloadable image generation and editing

Black Forest Labs released FLUX 3 Image Klein on 15 October. It is a distinct open-weight release from the FLUX 3 Image API covered earlier this month. The 5.6-billion-parameter generator accepts up to ten reference images and supports 4K output and bounding boxes for layout guidance. The release includes a base checkpoint for fine-tuning and a distilled checkpoint for faster generation.

A designer could use product references to position a bottle and props, then revise the background. Weights are downloadable, while BFL's API starts at $0.009 per 1K image. BFL says its INT8 build can run with 8 GB of NVIDIA GPU memory, but that is configuration-specific: the reference workflow needs more memory, and short prompts may use a separate upsampler. Bounding boxes guide placement rather than guarantee an exact mask. The FLUX Kommunity License governs commercial use; check the intended deployment and modification rights before building a service around it.

## Muse Spark Open: a large multimodal model for agents

Meta released Muse Spark Open on 15 October with 816 billion total parameters, 42 billion active parameters and a one-million-token context window. It takes text, images, video and audio as inputs and returns text. Meta positions it for coding and longer agent workflows, including a demonstration of footage review and vertical cropping in DaVinci Resolve.

A production team could evaluate it for finding relevant material across scripts, screenshots and edit notes. Meta's demonstration is not a XANTRANS test or a guarantee of reliable autonomous editing. Hosted access is listed through providers including OpenRouter, Together and Fireworks, with provider-specific prices. Weights are also offered for download. The highlighted local setup uses two M5 Ultra Mac Studios, a substantial hardware commitment. Hugging Face download access requests identity and organization information, and the Muse Model License carries commercial conditions and attribution requirements.

## Nemotron 3.5: visual documents and local agent work

NVIDIA released Nemotron 3.5 Super VL and Bolt 74B-A9B on 15 October. Super VL is a 120-billion-parameter vision-language model designed for agents working with webpages, charts, tables and PDF pages. A document-processing team could assess it for extracting figures from mixed charts and tables before passing them to a reviewer. Its model card supports up to one million tokens of context, although deployment settings can be lower, and limits video input to two minutes.

Bolt 74B-A9B is a separate model compressed for a 24 GB local NVIDIA system, aimed at agent tasks on RTX and DGX Spark hardware. Both models are available now as weights. Super VL also has hosted access and NVIDIA NIM deployment options. The smaller Bolt E4B 4GB is planned for 23 October and is not available as part of this launch. Hardware costs and serving setup remain part of local deployment; NVIDIA's speed comparisons are vendor measurements, not results from our own tests.

## Ultramarine 1 Flash: professional deliverables through an API

Paper Instruments published its Ultramarine 1 Flash report on 14 October. Built by post-training GLM-5.3 Flash, the model targets work such as financial analysis, consulting decks, legal timelines and operational documents. The report describes training tasks with workspaces and detailed grading rubrics, aiming at complete deliverables rather than only short answers.

A team preparing a client presentation could compare its draft against an existing process for accurate figures, source handling and layout. The company's benchmark wins and per-task savings come from its own evaluations, so they do not establish savings on a particular client's work. Listed API prices are $0.15 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens. Weights are described as coming soon. Despite the open-model positioning, access today is through the API, not a downloadable checkpoint.

## Sources

- [Pika Plugin for ChatGPT](https://pika.art/mcp/chatgpt) (checked 2026-10-17)
- [Black Forest Labs: FLUX 3 Image Klein, 15 October](https://bfl.ai/blog/flux-3-image-klein) (checked 2026-10-17)
- [FLUX 3 Image Klein base model card](https://huggingface.co/black-forest-labs/flux-3-image-klein-base) (checked 2026-10-17)
- [Meta: Muse Spark Open, 15 October](https://research.meta.ai/blog/muse-spark-open) (checked 2026-10-17)
- [Muse Spark Open model card and license](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) (checked 2026-10-17)
- [NVIDIA: Nemotron 3.5 Super VL and Bolt, 15 October](https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/) (checked 2026-10-17)
- [Nemotron 3.5 Super VL model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4) (checked 2026-10-17)
- [Paper Instruments: Ultramarine](https://www.paperinstruments.com/ultramarine) (checked 2026-10-17)
- [Ultramarine technical report, 14 October](https://www.paperinstruments.com/blog/ultramarine) (checked 2026-10-17)

Prepared with AI assistance from the linked primary sources. XANTRANS has not independently tested these releases.


Original page: https://xantrans.com/blog/ai-news-2026-10-17-pika-open-weight-models/