Cofounder & engineer. Running local AI, building things with it, and sharing what I learn.
Joined April 2022
- Tweets9.4K
- Following771
- Followers8.1K
- Likes16.3K
Pinned Tweet
Drama 3 went on a 4bit diet.
My unofficial NVFP4 quant of @fishaudio's Drama 3 drops 28% of its experts and peaks at 19.35 GiB instead of 36.95, with no measurable quality loss in my tests on one DGX Spark.
Capri, my local assistant, explains. Sound on 🔊
Links in the reply:
netrunner retweeted
Issue 3 of AGI Checkpoint is out 🎉
A few highlights for this week 🌟
・@MiniMax_AI put 428B M3.1 Flash weights up at $0.15/$0.60 per million tokens
・@Recursive_SI shipped Flint, a local coding model at
125+ tok/s on a MacBook Pro
・@plotarmordev got TensorFold cloning voices locally
・@ch0mpaa ran TensorFold on a Snapdragon
・@justinn_builds pushed REDSNOW to 275k context on two Sparks
・@Blackwellboy hit 2,008 tok/s prefill with BLACKFOLD
Follow these community builders, they are killing it 🔥
There might be a Kimi K3.1 Pro
A "Kimi K3.1 Pro tech blog" link showed up today in the Research menu on Kimi's site, right above the K3 blog, and then got pulled.
Makes me think Pro is a separate, higher tier above plain K3.1. No date, price or benchmarks yet.
This is one of the devices I'm most excited for this year.
I've always typed everything. But after reluctantly trying wisprflow, I'm actually getting used to speaking my prompts.
A ring with a mic and a trackpad, is the next step to level up my prompting game again!
Today we introduce OASIS 1.
The smart ring built for private dictation. Whisper to write. Touch to edit.
A first step beyond the keyboard toward a world where your intent follows you across every device.
Order at oasisdevices.com first batch is limited.
My iPhone Duo sniper has been running for about a week.
Today the alerts really picked up, but every time I got to one it was already gone. Starting to feel like scalpers have agents paying for them.
So I decided to give @Muse access to my Apple account, which has my card on file, so it should be able to pay for me.
I wanted to do it through @1Password new agentic setup that came out yesterday, but it's not working (on @bot), since it won't let me use my subscription without a US phone number.
Anyway I feel like I set this up too late, curious to see what happens now.
My iPhone Duo sniper has been running for around 21 hours.
Every ~70sec it checks 152 model and store combinations for in-store pickup on Oct 23.
Overnight alone that was 328 checks, close to 50k lookups.
Duo in stock: 0…
The sniper works fine. On the same run, 24 control items show up as in stock, so the data is live. There just aren't any Duos…
2025: everyone could make a website. Every single one was a waitlist page for an app that never shipped
2026: everyone can make a 3D game. It's a 3D model that jumps, but it's their model
2027: your mom is a hacker. She changed the wifi password without calling you
Really cool to see my Drama 3 repo as a dependency in this Jarvis project 🔥
Next up from me is a very high quality Drama 3 quant. It cuts the size in half and keeps all languages, so the voice takes up a lot less of the Spark/VRAM and allows to run more things alongside it!
LocalRouter AI assistant. One DGX Spark. Full duplex.
This actually feels like Jarvis in a box! Built on @ashxhart TensorFold
🧠 Qwen 27B
🗣️ Drama 3 with streaming (Thanks @plotarmordev )
🖼️ Qwen Image Turbo 2.1
🎬 MiniMax H3 video
👂 Whisper STT
~0.5s after I stop talking, it starts with its first word. Interrupt it anytime. Set the delay as you please for responses.
Normal models stay loaded and can be swapped as needed.
Want a video model up? Then Whisper and Drama will make space. Once the video is generated, start voice back up when the video is finished.
The goal is actually fully modular, so consider this a super early version of what it could be. MCP by default, but also ships with a simple UI you can extend as you please.
Thin by design, so everyone can make it their own.
What features do people want next? What about TP=2 or Mac Studio versions?
Finally, Kimi K3.1 looks close!
A model called Kibitz showed up on LMSYS Arena today and says it's Kimi. A screenshot from yesterday shows @Kimi_Moonshot Kimi Code tool already naming K3.1 in its model switch warning.
Last time, K3 sat on Arena as Kivine for about 30 hours before it launched in July. Nothing official from Moonshot yet, but if that pattern holds we're talking days.
Can a 163 MB speech model be a local WisprFlow? Bark 3 is as accurate as NVIDIA's Parakeet, which is about 15x bigger!
Similar thing I've been chasing with drama3, getting it to run on 20GB GPUs so more people can actually use it. This actually has potential to live anywhere.
Today we’re releasing Underdog Bark 3.0, the #1 frontier speech model
It's a tiny 163 MB download as accurate as NVIDIA’s Parakeet (15x bigger). It beats OpenAI's Whisper large-v3 (57x bigger)
It's fast- transcribes 1 min of audio in <0.25s on a Macbook. On recorded meetings (AMI), Bark 3.0 gets 7.82% of words wrong. Parakeet gets 9.42%. OpenAI's Whisper large-v3 gets 13.63% error rate
Bark runs on both your Laptop & Phone. It’s free and open source under CC-BY 4.0
Try private on-device meeting notes & dictation, powered by Bark, in Underdog today: underdog.ai/bark
netrunner retweeted
sparkDash 2.70 is out ⚡️
This is a MAJOR release.
New pages
- Cost compare: see what your fleet's token usage would cost on any hosted model, including cached-prefill savings
- Temps: daily and weekly GPU temperature graphs for every unit
- Metrics history per unit (48h and 30 days, opt-in), with a History panel on the unit page
Engines & hardware
- New engines: Strata & Flare (tok/s and benchmarks) and Radiance (live rates)
- Intel Arc GPU support
- AMD GPU telemetry
- GPU clock cap shown on each unit, and you can set it from Edit unit. If the unit has no passwordless sudo, the dashboard asks for the password, or you can run it on the Spark yourself
Benchmarks
- Decode bench: pick the temperature (default 0.7) and turn thinking on or off for prose. Both show on copied results and the share card
- Tool Eval Bench "Show card": a shareable image of a finished result
- Comparison card: tick two runs and get one image showing the difference
- Compare a run from one Spark with a run from another, plus an expandable scenario list
- MAX_ACTIVE_BENCH_JOBS setting to raise the fleet benchmark limit
Overview & monitoring
- Overview cards list every loaded vLLM model, plus CPU temp, CPU power and the engine's parallel slot count
- Token totals for the last 24h, 6h and 1h, with one row per unit and model
- Optional card showing inference request activity
- Alerts to ntfy, Discord, Slack or a webhook (opt-in), plus a tab badge and desktop notifications
Other
- "Run both" button in the Models panel
- "Back to Overview" link on every unit page
- Published container image
Link below.
We've entered a new era of AI where no console is safe anymore.
Somewhere along the way, maybe around GLM 5.3, everything started to change.
i jailbroke the nintendo switch 2
so i modded cinnamoroll, peter griffin, hatsune miku, and more into Mario Kart World, a switch 2 exclusive
now watch my friends and i get cooked by bots
learn more and get the exploit pepsipu.com/blog/2026-10-swi…
I looked into this and long context on Flint is almost free in memory!
Only 14 of its 49 layers keep KV for the whole sequence. The other 35 only look at the last 513 tokens. That works out to 28 KiB of cache per token, so 100K tokens of context fits in under 3GB.
Flint by @Recursive_SI released a model that at 4bit should come roughly at 20GB, so it sits comfortably on a 128/64GB Spark.
Recursive's own number is 150+ tok/s on a 64GB RTX Spark in NVFP4. Custom kernels account for 55% of that speed without touching the weights! Looking forward to test this along with drama on the same machine ~
Never name your agent after a time.
Mine was called Tomorrow, so everything it delegates shows up as "Tomorrow's message" and the other agents read it as "do this tomorrow" 😂
Fully automated procrastination…
I wrote a memory system for Tomorrow, the assistant you've heard me talk about, and it's the reason Tomorrow now works almost autonomously a lot of the time.
A few days in and it's going really well. What breaks for me is the interface. I've been running everything through Telegram with a single Hermes by @NousResearch, and managing multiple projects that way turned into a total mess. I thought it would be acceptable with some UX updates I added, but it's totally impossible.
I'm looking for something better. Now I’m considering mousebot or monocode. Anyone have experience with either?
Or anyone knows options that could give grok bot experience locally? I've also heard that Hermes is close to releasing a mobile app, would be nice if it has a @bot like system for agents but smh I doubt…
Flint by @Recursive_SI released a model that at 4bit should come roughly at 20GB, so it sits comfortably on a 128/64GB Spark.
Recursive's own number is 150+ tok/s on a 64GB RTX Spark in NVFP4. Custom kernels account for 55% of that speed without touching the weights! Looking forward to test this along with drama on the same machine ~
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.
Up to 43% more tokens per second on MoE models, just from better kernels.
Those kernels are only one part of each step. The model also spends time picking which experts to use and running attention, so a speedup this big showing up in overall speed is huge!
Quantization is easy. Fast quantization kernels for every format × every GPU × every workload are not.
Today we're open-sourcing Humming with Ant Group: one kernel library for all of it — up to 5x kernel speedups, +14–43% E2E MoE throughput.
Blog + benchmarks: vllm.ai/blog/2026-10-21-humm…
