Pinned Tweet
Updated my GH profile with a list of what I'm doing on llama.cpp π
Why? Because sometimes I forgot what I did...
Xuan-Son Nguyen retweeted
I've been building Llama with the llama.cpp team β a tiny menu bar app that runs open models on your Mac.
Here it is in 35 seconds π¦
Xuan-Son Nguyen retweeted
llama.cpp v0.8.0
github.com/ggml-org/llama.cpβ¦
Xuan-Son Nguyen retweeted
llama.cpp b11938 on M2 Pro 32 GiB (using ~3 GiB with the full 128K context) at ~110 t/s with c1 running with Pi!
llama serve --hf-repo LiquidAI/LFM3-VL-2B-GGUF:Q4_K_M --host 0.0.0.0 --port 8080 -fa on -np 1 --top-p 1.0 --min-p 0.0
Introducing π»π΅πΌπΉ: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> π»π΅πΌπΉ-π·.π½π±: text, 256k context window.
> π»π΅πΌπΉ-π
π»-πΈπ±: vision-language model with video-input support for spatial understanding
> π»π΅πΌπΉ-π°ππππ-πΈπ±: tool calls, voice chat, reasoning in 18 languages.
> π»π΅πΌπΉ-πΎπππ-πΈπ±: an omni model combines vision, audio, and language.
> ππ·-πΈπ± and ππ·-πΎπππ-πΈπ±: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> π»π΅πΌπΉ-π
π»-πΈπ± is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> π»π΅πΌπΉ-π·.π½π± is #1 on the Pareto frontier of device intelligence πππππππ.ππππππ.ππ
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ π§΅
Xuan-Son Nguyen retweeted
NEW: Liquid AI just released LFM3 on Hugging Face. The model family brings text, vision, audio, omni, and decision models together on a new shared backbone (~2B params).
It can even do real-time video captioning locally in your browser on WebGPU with π€ Transformers.js!
Xuan-Son Nguyen retweeted
Introducing π»π΅πΌπΉ: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> π»π΅πΌπΉ-π·.π½π±: text, 256k context window.
> π»π΅πΌπΉ-π
π»-πΈπ±: vision-language model with video-input support for spatial understanding
> π»π΅πΌπΉ-π°ππππ-πΈπ±: tool calls, voice chat, reasoning in 18 languages.
> π»π΅πΌπΉ-πΎπππ-πΈπ±: an omni model combines vision, audio, and language.
> ππ·-πΈπ± and ππ·-πΎπππ-πΈπ±: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> π»π΅πΌπΉ-π
π»-πΈπ± is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> π»π΅πΌπΉ-π·.π½π± is #1 on the Pareto frontier of device intelligence πππππππ.ππππππ.ππ
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ π§΅
OK so while everyone stop there and asking why the model behaves that way..
I gave it full gas to continue from where the trace left of
Quite franky, the model also confused why that URL was given.. Just like it should not be there at all?
Trace, raw tokens, & minimal reproduction script found here on @huggingface: huggingface.co/datasets/Coheβ¦
Friendly reminder: Weight are open, you can look at the logprob
The very next output token `https://` falls outside rank 20 even with top_k=20
Feels wrong, but idk, someone please tell!
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-bβ¦
OK so I want to clarify one misconception here: my project is NOT "just an USB forwarder"
And unlike CUPS, you can literally send print jobs (raw PDF) and run scanner from its built-in web server.
Nothing need to be installed, it's driver-free.
First write-up for my journey to RE the printer is up!
One of the main goals:
Own the cartridge and print head: no more ink counter, no more print head error, no more proprietary cartridges.
blog.ngxson.com/hack-my-prinβ¦
Xuan-Son Nguyen retweeted
I was the opening act for a family of ducks.
Thank you, Berlin. What a GTC.
Europe is building its AI ecosystem and economy, from AI factories to agents to robots.
DSX POD: Vera Rubin modular factories built, cabled and tested in the factory, then delivered ready to produce intelligence.
Agents: models, harnesses, sandboxes, knowledge and tools. NVIDIA Agent Toolkit brings the pieces together so companies control what their agents know, access and do. Β NVIDIA OpenShell enhances safety.
Physical AI: new customers across AI-RAN, robotaxis and robotics.
Personal AI: after four years building with Microsoft, RTX Spark brings personal agents to the Windows PC.
Thank you to President @EmmanuelMacron, Prime Minister @andyburnham and Chancellor @bundeskanzler for hosting me. Great to celebrate Sferical One with Swedenβs technology leaders. And thank you to President @TokayevKZ for the warm welcome.
And to our partners and NVIDIA teams: thank you for making it happen.
The keynote is here: youtu.be/rF1EsryCAwc
The same aftermarket black cartridge (right):
- Stock firmware refuses to recognise, show unclear error status β
- My firmware: ey yo what sup man π«‘π«‘
24 hrs later and we are printing the first colored sheet with fine details πͺπͺ
Cartridge protocols are all mapped out. The ink counter is however not reprogrammable. But still, nice to know!
Check out my other side project (still around the printer!)
nitter.cf/ngxson/status/21147539β¦
Printed the first "Hello world" from my open-source firmware ππ
It's reversed (pun intended), but we are close!
Another side project while I'm at this:
nitter.cf/ngxson/status/21147539β¦
