community @NousResearch · ex @KPMG
Open Source
Joined October 2021
- Tweets20.6K
- Following1.5K
- Followers22.2K
- Likes43.7K
Flint runs at 216 tokens per second on my RTX 5090.
I gave it a small coding job inside Hermes Agent: write fizzbuzz[.]py, run it, show me the output. it wrote the file, ran it with python3 and came back with the correct output.
however, it slows down as the context fills: 216 tok/s on an empty context, 152 once 32K tokens are in.
Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine.
• Best-in-class fast local coding
• Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark
• Built from scratch in under 4 months
Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.
25 new stories on the Hermes Agent use cases page this week:
- a drainlaying business sends its quotes and invoices from Telegram, on the job site or in the truck
- one IT person at a 165-person travel company runs about 80% of the daily work through Hermes
- a food and beverage company runs ops, finance and approvals on three Hermes profiles
- a mushroom grow tent on a Raspberry Pi, through two Lion's Mane harvests
395 real stories now!
hermes-agent.nousresearch.co…
if you work like me you like to see what is done, needs attention and what is going to be the next task in your to-do.
a new Hermes Agent plugin named 'Jot' keeps all your checklists and plans as notes beside the chat in Hermes Desktop, and your agent can tick off to-dos and add new ones when you ask.
I am going to use it heavily, specially since right now I am using a separated app for notes taking.
round 2 for the Bonsai files: this year's AIME and HMMT, 63 problems, 32k thinking budget, my RTX 5090.
>Ternary-Bonsai-2-27B PQ2_0, 7.2 GB: 46
>Qwen3.8-27B Q6_K, 22.9 GB: 47
>Ternary-Bonsai-2-27B PTQ1_0, 5.9 GB: 43
question by question, PQ2_0 solved 6 that Q6_K missed and missed 7 that Q6_K solved. a tie on 63.
on the answers that finished, Bonsai is right 95.4% of the time, Q6_K 96.1%.
Underdog-Saluki-27B is Qwen3.8-27B packed into a 7.9 GB file, out since 8 October. I ran it through my thinking board on my RTX 5090, next to four other files of the same model.
four-task average, thinking on, 16k token budget, 1,583 questions each:
- Qwen3.8-27B Q6_K, 22.9 GB: 89.7
- Qwen3.8-27B UD-IQ3_XXS, 11.9 GB: 89.5
- Ternary-Bonsai-2-27B PQ2_0, 7.2 GB: 88.7
- Ternary-Bonsai-2-27B PTQ1_0, 5.9 GB: 88.7
- Underdog-Saluki-27B IQ2-mix, 7.9 GB: 85.7
Saluki keeps 95.6% of the 6-bit score. the two Bonsai files keep 98.9%, and the 5.9 GB one is about 2 GB smaller than Saluki. Saluki is behind both of them on all four tasks, and it spends 1.5x to 1.8x the tokens per correct answer.
Saluki has one advantage: it loads on stock llama.cpp, and the Bonsai files need PrismML's fork. its model card also leads on tool calling, which this board does not test.
has anyone run Saluki and Bonsai side by side on tool calls?
Hermes Wingtips #108: hermes memory setup
Hermes Agent keeps its own memory in MEMORY.md and USER.md. a memory provider runs alongside it and adds things like semantic search, knowledge graphs and automatic fact extraction.
one command walks you through adding one:
`hermes memory setup`
pick a provider from the list, answer its setup questions and start a new session. your built-in memory stays on the whole time.
great for a long project, a research habit or an agent you talk to every day!
3D artists and game devs, this one is for you!
- text, one photo or several photos to 3D, then texturing, remeshing, UV unwrapping, rigging and animation, all from a chat with your Hermes Agent
- a finished model downloads as GLB, FBX, OBJ, STL or USDZ, at no extra credit cost
- the Hermes Desktop gallery keeps what each model was made from: the prompt, seed, settings and every step that came after
- for 3D printing: printability checks and repair, multicolor 3MF, splitting into printable parts and a slicer handoff
can't wait to see what you build with it
seeing the Microduck in Jensen's finale made my morning.
I am also using Hermes heavily to train mine, I can't wait to receive it and show you all tricks my little boy can do :)
Jensen closed his GTC Berlin 2026 keynote with this Hermes Agent demo as the finale.
He introduced it by saying the way to build an agent is to ask an agent to build it for you, so he installed and set up a customized Hermes in one prompt.
Then he used Hermes, running on @nvidia RTX Spark, to create a simulated training environment for the @huggingface Microduck. Hermes planned an obstacle course with Nemotron and had GPT-6 Astra build it using FreeCAD, Omniverse libraries, Blender, Newton, and Isaac Lab. Hermes trained thousands of simulated ducks on that course until the validated policy was ready to run on the real robot in the physical world.
At Nous, we believe open source is a practical tool as much as a philosophical ideal. It gives users the ability to build powerful, complex workflows chaining numerous local+cloud models, devices, and tools themselves instead of waiting for a company to enable them.
I gave Qwen3.8 27B one prompt and one hour with /goal in Hermes Agent on my RTX 5090, and it built Mont-Saint-Michel (yeah, I'm French) as a voxel diorama.
this is the model I run every day.
> the tide rises and falls around the mount on a one-minute loop
> pilgrims on the causeway, sheep on the salt meadows, gulls overhead
> T speeds up the tide, W brings a rain shower, N turns it to night with the windows lit
a video shows what a model builds when nobody steers it for an hour.
I think it stills need some tweaks and improvements but it is actually very fun to build.
reference image in the first comment.
great video if you like learning how things work: the Anatomy skill lets Hermes Agent explain an idea as an interactive machine figure.
youtu.be/dXKnb0TC7QY
witcheer retweeted
Krea 3 from @krea_ai is now available in Nous Portal
Fast mode does 2k edits in under 8 seconds, and Ultra is a higher quality tier outperforming Nano Banana Pro
It supports precise controls: draw a box to create a bounded editing region leaving the rest of an image unchanged, or specify a hex code to apply precise color
portal.nousresearch.com
today, Krea 3 makes it to the top of image arenas.
- "Fast" pushes the pareto curve of speed and quality.
- "Ultra" ranks top-4, only surpassed by OpenAI models.
the only company valued under $1T in the top-10!
more below 👇
Spark-X2.6-4B, SparkLLM's new 4B model built for on-device use, runs on one RTX 5090 with most of the card free, even at the 64K context Hermes Agent needs.
I ran the maker's own 4-bit GGUF on llama.cpp and served it with tools on. its first tool call came back clean.
> only 9 of its 36 layers read the whole context. the other 27 look at the last 512 tokens, so the 64K window costs very little memory.
> however, I've only measured speed and one tool call so far. a full Hermes Agent session on it is the next test.
to point your Hermes Agent at it: start llama-server with --jinja, then `hermes model`, Custom endpoint, http://localhost:8080/v1
model: huggingface.co/XHToken/Spark…
GGUF: huggingface.co/XHToken/Spark…
Hermes Agent with llama.cpp: hermes-agent.nousresearch.co…
