James Walker 🇬🇧🇺🇦

@ivanfioravanti DXG spark prefill. M5max 128. EXO. More AI than existed on the planet 3 years ago in my house

Another video to show the speed of MLX Qwen3-Coder-Next-5bit running on M5 Max with pi and mlx-lm server. In the middle of the video I speed up to show at the end the speed with larger context. More videos to come using LMStudio and others!

@0xSero $100 on the way. Love you man

1 day, the kindness of people knows no bounds.

@sudoingX Shared count?

how much VRAM do you have right now

@jeremymcs You mean it’s going to get better???

Just wait till v2 is finalized. #GameChanger $GSD

GSD for the win - it’s like being interviewed by staff engineer who then goes off and doesn’t come back till the build is done. @gsd_foundation have reinvented to coding paradigm for me. I can’t believe how much of my life used to spend in VSC not understanding how any of that shit worked.

I had to open this abomination today, holy fuck. How can anyone use this?????

@alexellisuk @claudeai 😂You can throttle the power with nvidia-smi -q -d POWER if you are worried. Max isabout 350 but they can spike at 450w briefly

@jxwalker I'm on a HX1200 but @claudeai is worried and thinks I need 1500w in there now.

NVLink installed and showing 14 GB/s bandwidth Next job: getting Qwen3.5 27B running across both 3090s without OOM (Codex is plugging away at it)

@official_taches @oguzhandilber3 DM me Lex if you want to chat about commercial use of GSD in a large GSI

@oguzhandilber3 Cheers for the shoutout brother! Is this using Termius? How does it work for you?

Far from my Mac but still I’m ‘getting shit done’ from my phone! Shoutout to @official_taches and all contributers for the GSD-2! Uzaktan mac terminalimden GSD ile projelerime devam ediyorum arkadaşlar. Bedava GSD ve 5 dolarlık Alibaba planı ile yapılamayacak iş yok 💯

All kinds of craziness. A low latency proxy and router in rust with an in-memory low latency ML pipeline. I have never done anything serious in rust before. I now have 110k lines of shippable code that every LLM and security auditor tells me is ”enterprise quality”. It’s like a superpower. I have not touched Claude code or cursor since GSD2 shipped.

@jxwalker @official_taches @oguzhandilber3 This is absolutely amazing @jxwalker What are you building right now?

@official_taches @oguzhandilber3 7x24 i am feeding it more work. It’s like a form of madness. Even when I am asleep I am dreaming of what @gsd_foundation is working on and when I can get my hands back on the keyboard.

@official_taches @oguzhandilber3 I have turned into a gsd junkie. Night and day. My gsd 2 is relentless. I am running tmux with two panes with tailscale and termius. GSD is building and I have a codex CLI window auditing and giving feedback and helping me keep GSD on a leash.

@gsd_foundation You guys rock GSD 2 is relentless. There is nothing like it for stopping context rot.

Amazing review from AI Labs 👏🏻 youtube.com/watch?v=uEit1o…

@MemoryReboot_ @sudoingX Have a look at llmfit - it will save you a load of time trying to figure out what will fit.

Bought my first rtx 3090 Ready to dive into this local LLM rabbit hole Shoutout for @sudoingX for this inspiration See you on the other side

@alexellisuk CUDA_VISIBLE_DEVICES=0,1 \ vllm serve Qwen/Qwen2.5-Coder-32B-Instruct-AWQ \ --tensor-parallel-size 2 \ --gpu-memory-utilization 0.93 \ --max-model-len 8192 \ --max-num-seqs 8 \ --dtype auto \ --port 8000

@jxwalker I'm using llama.cpp which will split layers but won't parallelise inference like you're thinking of.. vLLM would be best for that. If/when I get an NVLink I'll be testing out vLLM for sure. How are you using yours?

I realised that some of my issues with the 3090s may have been from running 1x and not 2x PCIe cables from the 1200W PSU What agentic tasks would you run with 48GB of VRAM?

@ZenMagnets @0xSero This is the lair of my GPUs. They live like students.

I take back my disparagement of @0xSero 's AI rig for the janky PVC trim holding his GPUs up. His GPUs are clearly living the good life. Sitting on nice hardwood floors, breathing fresh human air, eating wet food every day. My shit's in the basement gulag, chained by trip hazard wires, coils whimpering unheard. Ty @LLMJunky for helping my nubass build it.

@ivanfioravanti I am officially broke.

What about Qwen Image 2512 4 steps Lora? 🥇 M5 Max 11 seconds 🥈 M3 Ultra 19 seconds I bet M5 Ultra will be ~5 seconds 🔥

@ivanfioravanti Magic 60w 😂

@ivanfioravanti What is your power brick? I have an M5max coming but was going to reuse my 140w brick

What about M5 Max throttling... 👌 New 3nm process applied seems working well! 16" so far so good. I'm pushing it like crazy, it reached 200W of power consumption (how?), thermal state normal! High Power ~78º C as max Automatic ~91º C as max Curious about 14".

@sudoingX I bought the nvlink. Can’t share memory so still hit ooms but will have to test impacting tensor parallelism over it.

@jxwalker 2x 3090 NVLink is the one. 48GB unified. run 70B Q4 with no offloading or 27B dense with 500K+ context. NVLink is the last consumer card that supports it and you have it. use vLLM for tensor parallelism across both. the DGX Spark handles anything at full precision. you're set.

drop your GPU below. i'll tell you exactly what model and config to run on it. here's what i've tested and verified on real hardware: RTX 3060 12GB - Qwen 3.5 9B Q4 - 50 tok/s - 128K context RTX 3090 24GB - Qwen 3.5 27B Q4 - 35 tok/s - 300K context RTX 3090 24GB - Qwen 3.5 35B MoE Q4 - 112 tok/s - 262K context 2x RTX 3090 - Qwen3-Coder 80B Q4 - 46 tok/s - full VRAM all running llama.cpp with flash attention. every number is real. every config is tested. if your card isn't on this list drop it below and i'll tell you what fits.

@ivanfioravanti 16 inch. I was going for the 4 and thought I could spill to my NAS if I needed to but it was only an extra 1k for the extra 4t and the performance of the Mac nvme is insane. My NAS is spinners on a 2.5 gb network so no comparison

@jxwalker 8TB! WOW! I took 4TB. Is it 16" or 14"?

I'll keep testing M5 Max and posting here some results. I will then create some articles to wrap up the various things I'm discovering.