https://xitter.cf/t.co/YZtDAdykYf curiosity put to work. move fast without losing the details.
New Jersey
Joined February 2026
- Tweets101
- Following227
- Followers6
- Likes81
Pinned Tweet
Started some benchmarking local AI models on Intel's Arc Pro B65 (32GB): 3 models, 2 llama.cpp backends, 7 rounds. Decided to document the process.
attack the findings, question the data, roast it.
Looking for feedback to further improve this process.
“SEO is dead,” said everyone whose AI search strategy is one FAQ page.
Traditional organic search behaviors are changing amonst the user
AI search does not have a 1:1 attribution, but this is the opportunity to expand your value and level up. SEOs should be looking to increase their value by driving performance in other channels as well. have you taken control of that direct channel yet?
Tie yourself to goals and KPIs that have SEO at their core, but connect more directly to your business’s revenue goals. This may be to start leading more O&O initiative, social channels (YT, X, IG, etc)
heres the repo
github.com/snowopsdev/t3-thi…
have been playing around more with turning my old CarThing from Spotify into my dedicated @t3dotcodes notification thingy.
added in some useful full screen alerts for when a thread needs my attention.
actually liking this more than I thought I would.
definitely dont need to buy into those fancy "dedicated ai controllers"...just reuse old hardware and point your agent at it and let it rip.
tested four ways to run AI models on Intel's 32 GB Arc Pro B65. The fastest tool depends on the job:
• One person writing code: Intel's vLLM, 1.7× faster than llama.cpp
• 8 people at once: exl3xpu, 5.5× more total output
• A ~50-page prompt: exl3xpu, 2.5× sooner to start answering
• A huge 125B model on one card: Strata, 3.6× faster than llama.cpp
• Easiest to set up*: still llama.cpp
Every comparison is against llama.cpp at its fastest, with the same tests. Full numbers below.
*technically all are easy to setup when you just point your agent to the work 🤷♂️
have been playing around more with turning my old CarThing from Spotify into my dedicated @t3dotcodes notification thingy.
added in some useful full screen alerts for when a thread needs my attention.
actually liking this more than I thought I would.
definitely dont need to buy into those fancy "dedicated ai controllers"...just reuse old hardware and point your agent at it and let it rip.
Anyone remember the Spotify CarThing?
Well I decided this morning to recycle it to get a quick view on subs usage and t3 code info. @theo
Not exactly the most “useful” but reusing old hardware is fun.
Claude: 0% left, locked until Monday. Codex: 85% left. Weekly AI quotas are now a resource I have to manage, like disk space.
time to quantize
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio
Got access to the big sibling of the Muse model I've been benchmarking all week. 28B fit fine. 816B is a different conversation... haha. Stay Tuned!
Couldn't find benchmark for the Intel Arc Pro B65. Not one.
So it's going through my machine:
32 GB VRAM
3 models, ~30B each
llama.cpp on Vulkan, (Omarchy)
Results below as they come off the line 🧵
/Phase 5
New reference numbers for the Arc Pro B65 (llama.cpp SYCL, -ub 2048), 16K prefill / decode:
• Qwen3.8: 824 / 23.7 tok/s
• Gemma: 617 / 22.4 tok/s
• Muse: 821 / 25.6 tok/s
Every future driver update gets checked against these.
llama-server context checkpoints on Intel Arc, in an 8-turn chat: every setup reused its cache, but turning checkpoints off was faster in 6 of 6.
Worst case, Gemma 4 on Vulkan: 124 s per turn on, 5.5 s off.
Gemma 4 on Intel Arc, time to first word on a 16-token prompt:
• SYCL: 1.09 s
• Vulkan: 0.72 s
With 4,096 tokens: SYCL 7.7 s, Vulkan 11.1 s. Do you notice a one-second start?
/Phase 5
Phase 5 on the Arc Pro B65 closes four loose ends:
A) a new -ub 2048 baseline
B) why Gemma is slow on tiny prompts
C) does a multi-turn chat reuse its cache?
D) can Qwen stop overthinking a coding task?
/Phase 4
Drivers and runtimes update constantly, and some updates quietly cost you speed.
Phase 4 on the Arc Pro B65:
- 10-minute canary against a pinned baseline.
- Anything that moves more than 5% gets flagged.
Have you ever lost speed to an update?
