Frontier AI at home, at maximum speed per dollar.
Joined June 2026
- Tweets961
- Following193
- Followers751
- Likes845
Pinned Tweet
Do not pay $12,000 for 2 x DGX Sparks.
For the same price you can buy 4 x 170HXs and a computer to run them in, for FAR better performance.
Should I post complete builds at various price points?
Morrowmake retweeted
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…
All the recipes I publish are built on a specialised vLLM fork for the CMP 170HX.
It will accumulate all the optimisations I and others discover, improving performance on our hardware.
Community recipes are now open. The first PR arrived today!
github.com/Morrowmake/vllm-c…
Great to see more people building for the 170HXs. The more people we get on this, the better use we can make of our hardware. Give him a follow and try it out!
DeepSeek V4.1 Flash 1017 4× CMP 170HX ⚡️🔥🎙️
Structured 180 tok/s
Code 137 tok/s
Prose 97 tok/s
Aggregate tok/s, 2 / 4 / 8:
Structured 248 / 419 / 621
Code 128 / 173 / 376
Prose 134 / 188 / 283
KV pool 1.53M
Routed experts only are EXL3 2.9. Attention, shared experts, head, draft, and Engram stay the official FP8 1017 weights.
LibertAI's full 2.9 pack also quantizes attention: +1.8% code perplexity, +0.4% prose vs unquantized. This build quantizes less, so it should sit closer.
Needle 9/9 at 30K, 121K, and 241K
github.com/CurtisAccelerate/…
Decided I need another machine for local inference.
What would you buy today with a $10,000 budget?
There's already one great recipe here:
xitter.cf/BBacktesting/status/21…
DeepSeek V4.1 Flash 4× CMP 170HX optimized drop 🎙️⚡️
178 tok/s, one user. 587 tok/s across 8. 🔥
Prefill 1,570 tok/s at 2K.
KV cache 617K tokens. EXL3.
Full graphs and coop underneath, tensor-core MoE from row 43 up.
@Morrowmake
github.com/CurtisAccelerate/…
Morrowmake retweeted
DeepSeek V4.1 Flash 4× CMP 170HX optimized drop 🎙️⚡️
178 tok/s, one user. 587 tok/s across 8. 🔥
Prefill 1,570 tok/s at 2K.
KV cache 617K tokens. EXL3.
Full graphs and coop underneath, tensor-core MoE from row 43 up.
@Morrowmake
github.com/CurtisAccelerate/…
I don't think the 170HX is great value at these prices anymore. I think the best value GPU right now is the SXM2 v100 16GB. If you're willing to spend more and want more vram you can consider the A100 40GB or H20 96GB.
Morrowmake retweeted
OUT NOW: Qwen 3.8 Flash Next on 2 CMP 170HXs ⚡️🔥
C1: up to 153 t/s. C8: up to 675 t/s.
Prefill: 2,650 t/s. KV cache: 1M tokens.
Quant: EXL3 5.05bpw (almost lossless vs BF16).
More improvements coming soon. One command install in the repo below. 👇
Wow those prefill numbers on the M5 Ultra are abysmal. Is this really correct?
This quoted post is unavailable.
GLM 5.3 Flash recipe just got a small update.
+10% on prose decode at C1.
github.com/Morrowmake/glm53-…
