CTO @conste11ation | generating utility
Joined November 2015
- Tweets796
- Following200
- Followers1.9K
- Likes311
That makes sense. I was playing around with developing some EXL3 kernels last week focused on 27b initially but the implementation is still pretty rough. Maybe I'll just adopt yours and try to add optimizations. I'll make a PR if I find anything worth sharing.
Nice! We need a more active community branch for CMP 170hx. I’ll see what I can contribute.
On your Flash Next implementation, do you see lots of remaining upside with TP2? I’m seeing great results with PP2, like prefill speeds over 6,000 tok/s and decode speeds over 200 tok/s at c=1. I wonder if PP is actually more optimal given the PCIE constraints.
github.com/AlexBrandes/cmp17…
@seanphan Let me know if you’d be open to a PR on your repo. I believe this is the fastest public recipe for Qwen Flash Next on 2x 170hxs.
New recipe from the home lab on 2x CMP 170HXs 👨🍳
Qwen 3.8 Flash Next
Single stream: up to 226 tok/s
8 concurrent: up to 840 tok/s
Prefill: 6000+ tok/s
Running W4A16 quant - PP2 - PCIe Gen2 x4 links
These cards are putting in work.
github.com/AlexBrandes/cmp17…
We have over 300 paid models available also! Free models rotate every few weeks to give everyone a chance to try them out.
The state of GPU market right now.
Chinese bride gifts her partner a flagship GPU at wedding — he breaks down in tears
She handed him an ASUS ROG Astral RTX 5090 that costs around $12,000
Yep, free models have limited usage based on capacity. Throughput is the same as the paid models though.
GLM 5.3 Flash and Qwen Flash Next are also available for free on Constellation Gate at 1M context and we don’t train on your data!
constellationgate.ai/models
Awesome to see the Meta team contributing open weight models for the community.
I have to say though, kind of a confusing mix of benchmark results here 🤔
Alex Brandes ² retweeted
The best price for tokens is $0.
Qwen 3.8 Flash and GLM 5.3 Flash are free on Gate, in Gate Chat or with an API key. Free models rotate every few weeks, so try these two while they're still available.
Two models, free to try.
@Alibaba_Qwen 8 Flash and @Zai_org GLM 5.3 Flash are live in Gate Chat, with the full Constellation Gate treatment: every request security checked, trimmed for tokens and recorded.
Sign up, open Gate Chat, start chatting.
Heck yeah! I guess I picked up the wrong motherboard because I can’t fit my 4x in a single case at all.
Are you going open frame for the extra 4?
Wait until you guys catch up on CMP 170hx GPUs.
You can still get this much VRAM for under $8k on EBay.
Once you buy the hardware, tokens become a sunk cost. That changes your mindset from one of token scarcity to one of token abundance.
I have all of this compute, what should I build today?
You don’t have to spend $20k either.
Qwen Flash Next on Strata if you have a NVIDIA GPU. On Mac, Qwen3.8-27b via an MLX framework like MTPLX or MLX-Serve would be my recommendation even though it's a dense model, you can still get decode above 50 tok/s.
Alex Brandes ² retweeted
I was tired of paying for Suno every month just to get background music for my videos.
And the free route wasn't much better. Hours digging through "royalty-free" libraries, checking licences, hoping a track wouldn't get flagged three months later.
So I built my own way out.
@audioascode is an open source Python toolkit that lets AI agents write music.
Your agent writes the music as code: notes, tracks, instruments. The library renders it into real audio on your machine. The agent can listen back, revise it, and drop it straight into a video, a game or a presentation.
Every sound is synthesized from code. No samples, no recordings, nothing borrowed. There are 70 playable instruments today, from violin and timpani to marimba and drum machine.
The film below is the test. I had an agent score Audio as Code's own launch video: an orchestral medley of public-domain classics (Strauss, Beethoven, Grieg and the Dies irae), arranged in code and played entirely by the library. 27 instruments, 1,775 notes, with every cut timed to the score.
It won't replace a real orchestra. But for the music behind my videos, I don't need to rent it anymore.
Hello, world. We're Audio as Code 🎼
Music for videos shouldn't mean another AI subscription, or hours digging through "royalty-free" libraries hoping nothing gets flagged.
Audio as Code is an open-source Python library that lets your AI agent compose music the way it writes code.
Describe the vibe. Your agent (Claude Code, Codex, anything with a shell) writes an editable score, renders it on your machine, and keeps every note revisable. No samples. No subscriptions. No licence headaches.
🎻 70 instruments, all synthesized from code, from cello to 808s
🎧 Hear them: audioascode.com/instruments/
🕺 Ode to Joy as techno, disco and drum & bass: audioascode.com/electronic/
⭐ Code (MIT): github.com/joaoCarvalho1000/…
🌐 audioascode.com
The music in this video was written by an agent using Audio as Code.