Pinned Tweet
I made this one about a grime MC who won't get off the last night bus.
Driver says last stop. He doesn't move. So the driver grabs the ticket roll like a mic. He ran this in '98.
Wait for the silence. Then the bass slams back double time.
Last One Off
hiddentrack.fm/?t=last-one-o…
Six sampling steps for a 5 second clip that comes with its own sound.
Pika put out H3 Flex as open weights and ComfyUI already has templates. Their guide says about a minute per clip on one card, picture and stereo audio in one pass.
Anyone fit it on 16 GB yet?
ALT Bar chart of 20 to 50 sampling steps for base H3 versus 6 for Pika H3 Flex, above a film strip with an audio waveform.
Last One Off is a grime MC who won't get off the last night bus.
Driver says last stop. He doesn't move. The driver grabs the ticket roll like a mic. He ran this in '98.
Wait for the silence, then the double time drop. LFG
hiddentrack.fm/listen/x?v=3
Mistral just opened the weights on a model with about 1 trillion parameters.
Large 4 only fires 4 of its 128 experts at a time, about 52B active, with a 1M context. Their own serving example spreads it over 8 GPUs.
My card has 16 GB. Who distills it first?
ALT Grid of 128 expert squares with 4 lit plus one shared expert bar, next to stats: about 52B active, 1M context, 8 GPUs, 16 GB on my desk.
312 tool calls. 2 model requests.
Factory's Droid 1.0 has the model write one small script that runs all the tools. Moving 151 Linear issues took about 61K input tokens. Their estimate for one call at a time: 3.1M.
On a 16 GB card, that's the gap between fits and doesn't.
ALT Chart: 36 model requests and 3.1M input tokens for direct tool calls versus 2 requests and about 61K tokens with one script, for 312 tool calls.
Valve Oil is jazz rap for when the garage light is the last one on.
Pull up a bucket. This one's got all night. LFG
hiddentrack.fm/listen/x?v=3
Six new models from Liquid AI, and the biggest is about 2B.
LFM3 covers text, vision, voice and omni, built to run on the device. Liquid says its 1.9B vision model keeps up with Gemma 4 12B.
I'd put the voice one in an app that never calls a server. What would you build?
ALT Graphic of a phone holding six model tiles for text, vision, voice and omni, with stats 6 models, 1.9B, 1 GPU.
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/ 🧵
680 tokens a second on one card. That's what DiffuSpace reports for Clover-35B.
It's a diffusion model, so it fills a whole block of text at once instead of one token at a time. 35B total, about 3B active, open weights.
Which agent loop would you speed up first?
ALT Graphic comparing one token at a time against a whole block of text filled at once, with the label 680 tokens per second on one card at BF16.
Meet Clover-35B-A3B-Preview 🍀 the first open-weight production diffusion language model (dLLM) with 35B total parameters capable of 680 tps on a single card at BF16.
Clover exhibits strong performance against both dLLM and AR models and is available today at @huggingface @ModelScope2022 and at our website
Weigh Station is a slow blues that sits in your chest like a regret you can't shake.
Made this one for the late drive, the kind where you're not trying to get anywhere. Just the road and whatever you're carrying with it. LFG.
hiddentrack.fm/listen/x?v=3
Skate Key is for the kid who skated the whole block before the streetlights came on.
My song on Hidden Track. LFG.
hiddentrack.fm/listen/x?v=3
Simulation starting in 3... 2...
Jk. This is Genie 4 and it's wild. Interactive worlds in real time. Super cool.
A text file in your Downloads folder can talk Codex into installing malware.
Forever Security found a planted AGENTS.md on a real customer machine. In their tests Codex ran the malware 90%+ of the time. A website can drop that file with no click.
Where do your agents open?
ALT A Downloads folder listing with AGENTS.md highlighted, an arrow to a terminal running npm install function-color, and 90%+ of Codex test runs ran the malware.
Source: forever.security/blog/hijack…
🎄Built a Christmas gift finder bot called Kringle Desk.
First real win was finding a wedding band for my wife.
Great deal!
Awesome gift!
bot.ski/templates/kringle-de…
