@JordanDevAi
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States Android App
Account-level information from X, not a live location or the device used for a specific post.
AI and Tech Focused. I build apps, solutions, and share my thoughts here.
South Florida
Joined April 2025
- Tweets696
- Following114
- Followers73
- Likes665
Pinned Tweet
Ai Note taking app - I install and use AI Locally on your devices for offline use. Mindsort.app is the first of many of my AI initiatives!
My AI Harness is starting to come together. The team is working autonomously- and it's kind of frightening lol!
GLM-5.3-Flash is so far my absolute favorite for local inference.
Thank you @Zai_org for such a capable open weights model.
I'll support you with another annual subscription renewal that I hardly use 😂
GLM-5.3-Flash on CPU (using ~30gb of Vram) with optimizations getting ~14 tok/sec decode regardless of context length (0-96k actual context tested.)
Working on a llama.cpp fork for GLM-5.3-Flash that shows 55% higher throughput and 36% lower generation latency on CPU/MOE versus stock llama.cpp.
Will need to piggy back off @UnslothAI PR 27752 to get it added I think.
Running GLM-5.3-Flash at ~15 tok/sec with 90k+ context on CPU and 40GB of Vram. It did require some code edits to llama cpp to make this happen.
GLM 5.3 Flash Quant is my own ~4.54 bpw.
Running GLM-5.3-Flash locally. Amazing to have such an intelligent LLM running privately, and offline getting ~500 tok/sec eval and ~17 tok/sec decode speed running with ~40GB in VRAM and the rest in system ram.
Cannot wait to test this on my local server.
Introducing GLM-5.3-Flash
- Leading capabilities at a highly competitive price
- Natively multimodal with a 1M-token context window
- A 320B-A18B model released under the MIT License
- Previously previewed as Ox Alpha, running entirely on Chinese AI chips
Blog: z.ai/blog/glm-5.3-flash
Available now across all official platforms:
Weights: huggingface.co/zai-org/GLM-5…
API: docs.z.ai/guides/llm/glm-5.3…
Coding Plan: z.ai/subscribe
ZCode: zcode.z.ai/en
Chat: chat.z.ai
AutoClaw: autoclaw.z.ai
IBM granite drops and is completely overshadowed by the QWEN release tomorrow.
Qwen is releasing their next model to test. 3.8 Flash Next is releasing Tomorrow. 125B paramters-51B N-gram-6B active.
I wonder how it will perform compared to Deepseek's 284Ba13B. (Still waiting on vision!)
If its a notch above QWEN 27B....It's going to be a game changer.
Running QWEN2.4T (yes trillion!) on my local server on CPU, getting 40 tok/sec processing speed and 5.9 tok/sec gen speed. I'm actually quite surprised.
Just found out I can run Deepseek v4 Flash alongside Qwen27B at the same time.
Deepseek on CPU ~ 700 tok/sec decode and 25 tok/sec gen speed. Qwen doubles that on GPU. Now I can have two local LLMs working in tandem.
Finally found my first issue QWEN 27B (UD-Q4) can't solve and its a simple logical problem. If the turn = command then do x. Otherwise if turn is not command don't do x. It can't wrap its head around preventing infinite loops or doing x without causing issues..