View all articles
newsopinionno-hypeai

No Hype: Pika H3 Flex delivers 4K via upsampling

JG
Jacobo González Jaspe
|
No Hype: Pika H3 Flex delivers 4K via upsampling No Hype
This article is also available in Spanish:Sin humo: Pika lanza H3 Flex con pesos abiertos para vídeo 4K

Pika H3 Flex generates 768p video in seven seconds of computation according to lapaasvoice.com. Heretic, DeepSeek, Nvidia, Mistral, and the Dallas Fed provide the week’s data.

Pika H3 Flex brings 4K upsampling to open weights

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Pika released the H3 Flex open-weight video model on October 29, 2026, lapaasvoice.com. The model uses six generation steps instead of 50 by post-training MiniMax H3, lapaasvoice.com. It produces 768p video in approximately seven seconds of transformer computation, lapaasvoice.com. Pika also offers a hosted service to extend footage to 4K resolution, lapaasvoice.com. Native generation stays at 480p and 768p resolutions, lapaasvoice.com.

What we think. The open weights stop at 768p; the 4K is Pika’s hosted service, not the model. We read the download, not the headline, and the download says 768p. If your final render needs 4K, budget for that external service.

The good bit. The model accepts up to 12 reference files, including images, videos, and audio, lapaasvoice.com.

What you can do. You can check your GPU capabilities for local AI for developers using these weights.

Heretic 2.0 makes safety removal easy

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Heretic version 2.0.0 released on November 1, startupfortune.com. The project has 37,000 stars on GitHub and reached 13 million downloads, startupfortune.com. A journalist used the software to strip safety alignment from Llama 3.3 and Gemma 3 in under ten minutes, startupfortune.com. The creator reports the software has spawned more than 3,500 decensored models, startupfortune.com.

What we think. A journalist needed under ten minutes, which is less time than this column takes to read. Alignment on open weights is a default setting, not a lock. If you pull weights from a third party, check who published them before you trust them.

The good bit. The recent release included twenty contributors, eleven of whom were first-time contributors, startupfortune.com.

What you can do. You can assess the risk of Google Gemma 3: Open, Multimodal, and It Fits a Mac Mini if you run unaligned models.

DeepSeek releases super-agent prototype

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. DeepSeek published the dsh-superagent-prototype repository on October 31, startupfortune.com. This follows the August 13 release of the MIT-licensed deepseek-harness, startupfortune.com. Epoch AI’s Capabilities Index shows DeepSeek, Qwen, and Kimi closing a 17.5 point MMLU gap from 2023, startupfortune.com. DeepSeek describes the new prototype as reference code rather than a finished product, startupfortune.com.

What we think. MIT licence plus reference code: we read it as study material, not something to deploy on Monday, and DeepSeek says so itself. If you try it, keep it in your development environment.

The good bit. The architecture is released under an MIT license, startupfortune.com.

What you can do. You can check what fits your own machine in Best local LLMs: which one to use for your hardware and task.

US Justice Department probes Nvidia acquisition

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. The US Justice Department is questioning Nvidia’s proposed purchase of Hugging Face, startupfortune.com. Nvidia announced the $12.9 billion deal on September 3, 2026, startupfortune.com. The Hugging Face platform hosts 3 million models and 200,000 companies, startupfortune.com. Nvidia is the largest contributor to the platform with over 500 models, startupfortune.com.

What we think. Three million models owned by the company that sells the GPUs is a lot of eggs in one basket, which is why the Justice Department is asking. If you depend on Hugging Face, keep a local copy of the models you actually use.

The good bit. Nvidia says the deal is on track to close in the first half of 2027, startupfortune.com.

What you can do. You can use a VRAM calculator to plan for larger models from the platform.

Mistral Large 4 targets cybersecurity

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Mistral AI released the open weights for Mistral Large 4, technews.bg. The model rejected 96.6% of malicious queries in specific benchmarks, technews.bg. Mistral’s API costs $0.70 per million input tokens and $2.10 per million output tokens, technews.bg. The model achieved a 51% score in the Artificial Analysis Cyber Index, technews.bg.

What we think. A 96.6% refusal rate on a benchmark is a lab result, not a guarantee against prompt injection. If you deploy it, throw your own attacks at it before you trust it.

The good bit. The model is designed to handle cybersecurity-related usage in real traffic, technews.bg.

What you can do. You can learn How to Classify Your AI’s Risk Level Under the EU AI Act (Art. 6) for your business.

Data centres drive up electricity prices

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. A Dallas Fed working paper estimated US wholesale electricity prices rose by 3% to 5% in 2024 and 2025 due to data centres, abacusnews.com. An MIT working paper found that residential electricity rates rose by about 2.1% after a data centre entered a utility area, abacusnews.com. Some regions could see a 50% increase in wholesale prices through 2028, abacusnews.com.

What we think. Data centres now raise the bill for people who live near them, not just for the people who use them: about 2.1% on residential rates. If you are costing a local server, put the electricity in the sum.

The good bit. The findings are based on specific modeled scenarios and working papers, abacusnews.com.

What you can do. You can check hardware by job to see if local hosting remains affordable.

In one line

Pika H3 Flex provides fast 768p generation via open weights.

Work with us

If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.

Diagram
Share: LinkedIn X
Veredicto semanal

Get new guides before anyone else

Subscribe and we tell you when new guides, templates and workflows go up. One email a week, no spam.

Already published: 102 guides and 25 templates. All free, no signup.

Bonus: the local-AI starter pack PDF when you subscribe
Once a week No spam Unsubscribe anytime

See what you get