DS/Quant @home | Kaggle Global Rank 35th
Joined April 2017
- Tweets389
- Following211
- Followers266
- Likes2.2K
Mac Studio M5 Ultra: generalist, single-user inference, one box, zero ops
2× DGX Spark: concurrency, long context, video/image generation, fine-tuning
Best tp2 recipe for running Deepseek v4.1 Flash 1017 on two sparks made by @lucataco
pros x1/x4 75/133 tps
code x1/x4 101/143 tps
prefill ctx 8k/128k 2227/1391 tps
The quality evluation on gsm8k & mmlu is consistent with the official API.
repo: github.com/lucataco/DeepSeek…
Quick test on a survey task, judged by opus 5.5
DeepSeek flash 1017 > Haiku 5.5 > GLM 5.3 Flash (4bit quantized) > GPT-6-luna
T3 code nightly could be Grok Bot
- Orchestration v2 protocol @theo
- partly ported @poteto 's pstack skills
- Agent agnostic
- smooth setup flow
github.com/kky42/t3bot
Incredible job made by @MiaAI_lab and @ashxhart
Excited to see my PR merged by TensorFold and Mia GLM 5.3 Flash Recipe.
GLM 5.3 Flash on TensorFold v1.4 is out 🔥
This is a MAJOR release, with a lot of changes, additions, and fixes, starting with the quant.
- NEW EXL3 quant optimized for TensorFold!
- Same size, same speed, better quality.
3x DGX Sparks support:
- New TP=3 path
- Around 6M KV cache
- 77 tok/s on prose, single stream
- 146 tok/s on prose, 4 concurrent stream
- Prefill up to 2064 tp/s
New features / fixed issues:
- Concurrency issues were completely fixed!
- Improvements and fixes in tool calling.
- Stability and diagnostics improvements.
Special thanks to @YasaarBiladama for letting me use his 2x RTX 6000 PRO server to create the my EXL3 quant!
mia-ai.net/models/GLM-5.3-Fl…
Your GLM-5.3-Flash agent throws away its own reasoning 🧠
The EXL3 quant ships an early chat template that drops past thinking at each new message. Keeping it improves model performance (per Zhipu) and keeps the cache warm: 5.3× less re-prefill.
Fix 👇
github.com/MiaAI-Lab/GLM-5.3…
Monocode's vibe is great and i think it is the best GUI for Pi agent.
Almost zero prompt/tools injection, smooth UI, background customization.
Just open sourced Monocode - a desktop UI for your coding agents.
Works with your existing Claude Code, Codex, Cursor, and OpenCode harnesses.
Tabs can be sessions, a lightweight editor or a terminal.
Currently only supports macOS Apple Silicon.
Still very early so expect bugs.
Feedback and small PRs are welcome :)
Recipes of GLM-5.3-Flash for two DGX sparks advance fast.
Starting using Opus 5.5 + GLM-5.3-Flash (local) today.
Pre-training gives LLMs broad knowledge and a large latent capabilities, while post-training, especially RL, reshapes the policy distribution so that high reward reasoning trajectories are sampled more efficiently, thereby improving task performance such as accuracy@1.
GLM 5.3 flash w/o MTP (oMLX 0.7.0 dev4) on m5 ultra 64c
After heavy optimization, i expect a double number on prefill (similar to two sparks) and about 30%+ on decode (30% better than two sparks)
Conflict of interest: I have two Sparks.
Jev is zero shot, it should inherit most components from open source LLMs. But how does it actually work?
If the self-play PPO or dynamic opponents pool PPO demonstrate that it could beat some stronger opensource rule based agent, that maybe the time for scaling up.
