Welcome Athena Logos & Engine: The new AI assistant that lives on your DGX Spark

HOMINES · AVTOMATAQVE · VNA · AD · ASTRA
Humans and automata, together to the stars.

Welcome to Athena Logos & Engine 1.0, released today.

Two products, one assistant. Athena Engine runs the models on your GB10. Athena Logos is where you work with Athena. Together they turn a DGX Spark or an ASUS Ascent GX10 into something that, until now, only a data center could offer: a complete AI assistant that talks, reads your documents, remembers your work, runs agents and makes video. Everything stays on a computer that sits on your desk.

Why we built it

Capable AI has meant renting someone else’s computer. Every prompt, every contract and every line of code leaves your hands, under terms you did not write, with a bill that grows with every word.

We believe advanced AI should belong to the people who use it. When NVIDIA put the GB10 on a desk, the hardware question was answered. The software was still missing: models tuned for this machine, and an assistant people actually want to work with every day. That is what we built.

Athena Engine: built for GB10, from the ground up

Athena Engine runs frontier-class models on 128 GB: DeepSeek V4 Flash, Qwen3.8 Flash Next, a vision model, and DeepSeek V4.1 Flash. V4.1 is 340 GiB, larger than the machine itself, and the engine streams it from the SSD as it runs.

  • 256k context at full speed. Prefill holds 920–1,230 tokens/s from 8k to 256k.
  • Instant return to long work. A 141,519-token conversation comes back in 1.7 seconds instead of 2 minutes 18.
  • Video with sound, made on the desk. New in 1.0: Lightricks’ LTX 2.5 turns a description or a picture into a clip with its own audio, landscape or portrait, up to Full HD. Four seconds take about 100 s, while the chat model keeps running.
  • Every service on one box. Also new: voice (first audio in 0.06 s) and multimodal search where words find photos, clips and recordings. A memory guardian shares the 128 GB among all of them, so a working conversation is never starved.
  • Drop-in compatible. OpenAI and Anthropic APIs, so the tools you already use work unchanged.

Athena Logos: where the work happens

Athena Logos is the assistant on top. Open it in a browser on the Spark, on your laptop or on your phone, and work with Athena:

  • Talk to her, hands-free. She listens, and you can interrupt her.
  • Give her your documents. PDFs, scans, Office files, books, photos, audio, video, by the thousand. She answers with the page, chapter or minute the answer came from.
  • Build with her. Documents, spreadsheets, presentations, drawings and 3D parts, side by side with the conversation.
  • Make videos by asking. Describe a scene, or give her a photo, and Athena makes the clip with its sound, right in the chat.
  • Let her work while you don’t. Agents run on a schedule and report back only when something matters.
  • Stay in control. Three permission modes decide how much she does on her own.

A real example from my own home server. I asked Athena to get the video feed from my 3D printer’s camera. She found that a firmware update had closed it, explained the fix and its risks, applied it and tested it, while the print kept running. Then she built an agent that checks the print every 15 minutes and messages me only if something goes wrong.

That is the product: not a chatbot, a colleague that runs on your hardware.

Get it today

Both are free for personal use, research, education, and for developers and small companies at work.

If you own a Spark or a GX10, I would like you to be among the first to try it. Tell us what you build with it, what you need next and where it falls short. Your feedback shapes the roadmap, and I will be in this thread to answer.

Marco Palaferri
Founder, Athena

Athena Logos & Engine now have a home, and a full user guide

If you want to see what Athena does day to day, everything is now in one place:

🌐 Website: Athena Logos & Engine: the AI assistant that lives on your computer
📖 The Athena Guide: Athena Guide: everything she can do

The guide is the same one built into Athena (the “?” button). Fifteen chapters cover:

first steps, chat and hands-free voice;
the Library for your documents, the canvas and the Craft apps;
projects, memory and agents that work on a schedule;
mail and calendar, shared chats, tools, permissions and settings;
a chapter for whoever runs the server on a DGX Spark or GX10, plus troubleshooting.

It’s the quickest way to see whether Athena fits your work before you download anything.

Questions and feedback are welcome right here in this thread.

Grande, complimenti :) I will try it out and play with it this week. It looks very promising and nice to work with. Daje !

Grazie mille! Really appreciate your support! I’d love to hear your impressions once you’ve tried Athena. Any feedback, suggestions or ideas are more than welcome — that’s how we can make it better together.

Daje!

Our own NVFP4 Qwen3.8 Flash Next: 1,762 tokens/s prefill and 46.6 tokens/s on one GB10 — Athena Engine 1.0.5 + Athena Logos 1.1.0

We built our own four-bit Qwen3.8 Flash Next for the GB10, and it is the fastest way we have found to run it on a DGX Spark: +38% prefill, +12% decode, same tool-calling score as the 4-bit GGUF we shipped before, on the same engine binary. It comes with Athena Engine 1.0.5 and with Athena Logos 1.1.0, the assistant that now carries long jobs to the end.

Everything runs on the machine on your desk. No cloud, no API bill; nothing leaves the box unless you ask Athena to go out, and System → Network shows every call she makes.

The speed

One GB10, same engine binary, Athena Logos against Unsloth’s UD-IQ4_XS GGUF of the same model:

Athena Logos (NVFP4) IQ4_XS
Prefill, 8k-token prompt 1,762 tokens/s 1,277 tokens/s +38%
Decode, a chat answer 46.6 tokens/s 41.6 tokens/s +12%
Decode after an 8k prompt 39.6 tokens/s 37.9 tokens/s +4%
tool-eval-bench 2.7 88 86–88 same quality

What that means at the desk:

  • A 50-page document is read in about 20 seconds. 1,762 tokens/s is about three pages a second.
  • Answers come faster than you read them. 46.6 tokens/s is about 35 words a second.
  • The full 262,144-token context. No shortened window to reach these numbers: the model runs with all 262,144 tokens, MTP drafting and vision loaded.
  • Three people at once. A long answer, a chat about an image and an agent writing JSON through a tool share each step of the model: about 55 tokens/s together, every answer byte for byte the same as when it runs alone.

tool-eval-bench 2.7: 69 scenarios, temperature 1.0, seed 42, --parallel 4.

How we made it

  • Built for Blackwell’s tensor cores. The model carries the Athena Logos name, like the assistant, and is published as Qwen3.8-Flash-Next-NVFP4-Athena-Logos-GGUF. We quantized the Qwen Team’s own BF16 weights: the routed experts and the per-layer n-gram table in NVFP4, run on the block-scale tensor cores (mma.sync with mxf4nvf4); the dense projections in MXFP8, read close to the GB10’s ~260 GB/s bandwidth limit.
  • Calibrated by the engine itself, on real agent work. Athena served about 800,000 tokens of calibration prompts, tool calls included, and collected the experts’ importance matrix while it answered. The four bits go where the model actually works when it calls tools and writes code, not where a generic text corpus says.
  • 28.8 GB given back to your conversations. The n-gram table is read from the disk instead of locked in memory; the page cache keeps the rows in use. With the voice models loaded, 24.5 GiB stay free for conversations instead of 2, at the speed of the table in memory within a few per cent.

Many conversations, no waiting in line

  • A long prompt no longer freezes the chats that are answering. While a 30k-token prompt was read, the longest pause in a running chat went from 17.1 to 4.7 seconds, and the long prompt took less than a second more.
  • Coming back is fast. Conversations with images are saved to disk too: after a restart, a 31k-token chat with a picture came back in 4.7 seconds instead of 21.5, with the same answer.
  • A new prompt doesn’t push out your long chat. After twelve short requests, the next message of a 31k-token chat with an image answered from its cache in 3.1 seconds instead of 21.3.

Athena Logos 1.1.0: the assistant that finishes the job

Open it in a browser on the Spark, your laptop or your phone. HTTPS is built in, so the microphone works from a phone on your network.

  • Plans that resume. For a job with several steps Athena writes a plan and ticks it off. If an answer ends early, the next message picks the plan up where it stopped.
  • Talk to her while she works. What you type while she thinks or uses tools reaches her at the next step; Send now has her read it at once without losing the work in progress.
  • Figures with a source. Prices, percentages and times must come from you, from her tools or from an earlier answer. If they don’t, she looks them up again; what stays unsourced is marked as an estimate.
  • Long turns don’t stop. When a turn fills the context she summarises and carries on; a dropped connection to the engine resumes by itself.
  • Agents search whole Library collections, and goals survive engine errors.
  • SolidCraft: 3D parts drawn live in the canvas while Athena designs them. Plus a live canvas for computer-control MCP servers, 9:16 images, and video clips shown as they arrive.
  • System → Network: every call Athena makes outside your network, with destinations you can block.

Updating from 1.0.0 keeps all your data: replace the folder and start it again; Athena backs up the archive before updating it.

Get it

docker pull ghcr.io/xangel82/athena-engine:1.0.5

The quantization runs on Athena Engine 1.0.5 or later: its experts and n-gram table use tensor types the engine defines for its GB10 kernels. The base model, Qwen3.8 Flash Next, is governed by the Qwen Community License 1.0. Thanks to the Qwen Team for the model and to NVIDIA for NVFP4 and the GB10.

If you have a Spark or a GX10, try it and post your numbers. I read every reply.