Local Inference maniac, now I spend more time exploring Agentic Engineering than gaming
0xMX
Joined February 2009
- Tweets48.3K
- Following1K
- Followers895
- Likes75.4K
Glitch 81 ᯅ retweeted
LFM3-VL-2B vs Gemma 4 E2B on 50 real scanned receipts.
LFM3 got 41 of 50 receipts fully right. Gemma got 28. 50 receipts from SROIE, same prompt, JSON out, one 5060 Ti 16 GB.
- LFM3-VL-2B: 95% of fields right, 41/50 receipts perfect, 46 s
- Gemma 4 E2B: 87% of fields right, 28/50 receipts perfect, 62 s
Across both tests: Gemma reads messy chat-style text a bit better (95% vs 90%). LFM3 reads receipts better, and was faster both times.
@liquidai @googlegemma
Introducing 𝙻𝙵𝙼𝟹: 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.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: 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.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 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.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
Glitch 81 ᯅ retweeted
CUDA Live: Introduction to CUDA Rust xitter.gq/i/broadcasts/1RKZzBapW…
Glitch 81 ᯅ retweeted
Want more tokens per second on your NVIDIA DGX Spark? ⚡
Explore three techniques to accelerate MoE inference, from optimized kernels and FP8 quantization to multi-token prediction.
ALT Rectangular device with textured front on a wooden table, greenery in background.
Glitch 81 ᯅ retweeted
Silicon Valley has officially gone too far. 😂
First, we couldn’t stop checking our phones at the gym, now we’re checking whether our AI agents have finished the code between sets.
New gym rule: No vibe coding between sets.
Forget leg day. We’re skipping debug day.
Remember: Your AI agent can wait. Your gains can’t.
Although, knowing tech bros, someone’s already building an AI personal trainer that lifts for them. 🤣
Glitch 81 ᯅ retweeted
‼️ ALERT - Hackers exploited a critical Cisco SD-WAN flaw (CVSS 10.0) to plant "KeyDrift" malware designed to steal encryption keys protecting traffic between sites.
The malware also provides backdoor access, with infrastructure links to a network used by Volt Typhoon.
Full details ↓ thehackernews.com/2026/10/ha…
Glitch 81 ᯅ retweeted
New in Claude for Legal: a gallery of ready-made artifacts for contract review, regulatory horizon scans, cite checks, and more.
Start from the built-in example, add your own documents and connectors, and share it with anyone across the business.
Glitch 81 ᯅ retweeted
⚡️ Your Qwen3.8-27B GGUF already ships a free draft head.
People thought Ollama was ~2× faster than llama.cpp on the same file.
It wasn't the engine — Ollama just turns MTP on by default (n=4). Stock llama.cpp leaves it off.
One RTX 5090, Q4_K_M:
off → 73.6 tok/s
MTP n=3 → 133.6 tok/s (1.81×)
quality stayed at 100% (speculative = lossless)
Flag: --spec-type draft-mtp --spec-draft-n-max 2
(with q4_0 KV they hit 136.7 tok/s around 17 GB)
One card, one build — your mileage will move. Still: if you're on stock flags, you're leaving free speed on the table.
via laxmimerit 👇
huggingface.co/Qwen/Qwen3.8-…
Glitch 81 ᯅ retweeted
In the spirit of transparency and openness, we are releasing the traces so developers can test this out for themselves. We noticed this behaviour without prefilling with the following prompt, but this is rare. If you want to reproduce it, you can use this prompt with prefilling.
Open models make for better science and are part of why this problem came to light at all: developers running Qwen3.8 on their own machines spotted the strange links, and having the weights let us study the behaviour in depth.
We're encouraged by the thoughtful discussion our research has sparked across the AI and security communities. It's our belief that open-weight models are an important part of the future of AI, and we've recently released several under the Apache 2.0 license.
Trace, raw tokens, & minimal reproduction script: huggingface.co/datasets/Cohe…
Glitch 81 ᯅ retweeted
What the hell…
You can play 3rd Strike & MvC2 in a Browser now?
streetfighteranniversary.onl…
marvelvscapcom2.online/
Glitch 81 ᯅ retweeted
Cohere knew their horse shit models couldn’t compete with QWEN. So they decided to spin a narrative.
Go on, publish the full logs and traces if your so confident.
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-b…
Glitch 81 ᯅ retweeted
BREAKING: Microsoft, $MSFT, is offering buyouts to 60,000 workers as part of the company’s efforts to cut costs amid AI adoption, per Bloomberg.
Personnel on select teams in the US, Canada, India, and the UK at a vice president level and below will be eligible.
This reflects ~27% of Microsoft's total workforce.
Glitch 81 ᯅ retweeted
UPDATE: ra-jit got renamed to Blaze!
Glitch 81 ᯅ retweeted
You can convert any USB-only printer to AirPrint using a cheap ESP32-S3 😉
No Raspberry Pi required!
Glitch 81 ᯅ retweeted
Simplest explanation is just a narrow RL training reliant on their internal infra (which isn't good design for an open weight model)
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-b…