Antonio Martinez

GenAI Engineer @ NVIDIA

[Antonio Martinez @antoniomtz]

@tonbistudio https://t.co/JBDKxyUQd1

[Antonio Martinez @antoniomtz]

@Teknium Staying busy is the best therapy. Keep up the good work @Teknium

[Antonio Martinez @antoniomtz]

@Teknium Sorry to hear that. I went through the same last year. My dog passed away after 11 years with me.

[Satya Nadella @satyanadella]

https://t.co/xv6csf1SbV

[Antonio Martinez @antoniomtz]

@Teknium @Teknium any updates on this?

[Antonio Martinez @antoniomtz]

A new toy has arrived ! Can’t wait to test it with local models on the DGX Spark.

@huggingface #reachy #huggingface #dgxspark

[Teknium 🪽 @Teknium]

Had Hermes Agent with the Manim Video skill plus it's TTS tool create a video explaining Hermes' Agent.

[Satya Nadella @satyanadella]

https://t.co/vLmiBKTtX3

[Antonio Martinez @antoniomtz]

The best combo.

#hermesagent #hermes #nvidia #dgxspark #AI #nousresearch

[NVIDIA AI @NVIDIAAI]

Love seeing photos like this.

Let's see some more setups.

Curious what everyone's running these days. What models are you using and what are you building?

[Antonio Martinez @antoniomtz]

@steipete @steipete would you use VISION.md + /goal ?

[Antonio Martinez @antoniomtz]

This.

[Naval @naval]

Software platforms are going to be rebuilt for agent-first.

[Antonio Martinez @antoniomtz]

Nvidia skills are now part of Hermes agent.

[Nous Research @NousResearch]

Full Release Notes: https://t.co/HNnmi3QBmy

Hermes Desktop: https://t.co/PP4eGRXATZ

To update, run: 'hermes update'

[Antonio Martinez @antoniomtz]

I’ve had a few "wow" moments over the last few months working with agentic AI harnesses.

This weekend I had another one.

I ran Hermes Agent on an NVIDIA DGX Spark and gave it a single prompt:

/goal Follow the NVIDIA tutorial to fine-tune a FLUX model locally, troubleshoot issues autonomously, resume long-running tasks automatically, and validate the final model by generating an image.
Tutorial: https://t.co/IMoDbrvltC

For the next ~4 hours, Hermes agent autonomously:

- downloaded and validated ~30GB+ of FLUX model weights
- built Dockerized training/inference environments
- launched and monitored 100-epoch DreamBooth LoRA fine-tuning
- created its own recovery/resume scripts
- scheduled cron-based orchestration for long-running tasks
- handled troubleshooting without supervision
- wrote Python scripts to validate the workflow through the ComfyUI API
- generated the final fine-tuned image successfully

And it completed everything in one shot!

One thing that stood out to me was how capable DGX Spark is for local AI customization workflows.

Being able to run DreamBooth LoRA fine-tuning locally on a compact AI supercomputer changes the equation for teams that want to customize foundation models for their own business requirements, proprietary datasets, products, or domain-specific use cases.

The stack was:
- NVIDIA DGX Spark
- FLUX.1-dev (12B model)
- DreamBooth LoRA fine-tuning
- ComfyUI
- Hermes Agent
- GPT-5.5 reasoning/planning

If you want to get started with DGX Spark, NVIDIA created a bunch of recipes and tutorials for different use cases: https://t.co/64aVChEhqf Just point your AI harness into the tutorial and let it do itself.

@NousResearch @NVIDIAAIDev @bfl_ai #hermes #nvidia #dgxspark #flux #Nous

[Antonio Martinez @antoniomtz]

@atomic_chat_hq @atomic_chat_hq could you please share the prompts?

[Antonio Martinez @antoniomtz]

@Angaisb_ @Teknium Yes

[Antonio Martinez @antoniomtz]

@Teknium @Teknium can you share the best practices to upload your "agent" as a github repo? this is a very important feature.

[Antonio Martinez @antoniomtz]

@HermesAgentTips Linux

[Antonio Martinez @antoniomtz]

@tobi @tobi it would be great to share your workflow for the autoresearch. I'm doing similar work with Hermes + Qwen on DGX Spark.

[Antonio Martinez @antoniomtz]

@HermesAgentTips I have found Qwen3-Coder-Next-FP8 to be more reliable and accurate with decent throughput. (Tested on DGX Spark)