Running Kimi K2.5 on my desk.
Runs at 24 tok/sec with 2 x 512GB M3 Ultra Mac Studios connected with Thunderbolt 5 (RDMA) using @exolabs / MLX backend.
Yes, it can run clawdbot.
Running Kimi K3 on my desk?
It will require 4 x 512GB M3 Ultra Mac Studios (2TB @ 3.2TB/s).
Waiting on active parameter count, but based on the rumours I expect we can run it at 30+ tok/sec with MTP + Tensor Parallelism using RDMA over Thunderbolt 5.
Prefill will be slow but 512GB M5 Ultra should make that ~5x faster (expecting in October).
It will be on local dot ai with full benchmarks as soon as the weights drop. Comment below for early access - sending access codes out throughout the day.
Running Tencent Hy3, a 295B-parameter MoE, on one RTX 4060 Ti 16GB.
How: NeutronStar (my CUDA fork of @antirez's ds4) keeps attention + shared experts resident and streams the routed experts off an SSD per token, so a 295B model runs on a $400 card. Had to write a new GQA attention path (ds4 was MLA-only) and build a ds4-native 2-bit GGUF.
~1.8 tok/s, interactive chat. Weights:
https://t.co/IWsxg6qqYm
GLM 5.2 @ 19.8 tok/sec on 2 DGX Sparks ⚡️
I switched to a dspark draft model with K=2 (model from @RedHat_AI)
Acceptance is ~68% because the drafter was trained with the full FP8 model...
Next step is to fine tune the drafter so acceptance on this super quant goes up
Clustering NVIDIA DGX Spark + M3 Ultra Mac Studio for 4x faster LLM inference.
DGX Spark: 128GB @ 273GB/s, 100 TFLOPS (fp16), $3,999
M3 Ultra: 256GB @ 819GB/s, 26 TFLOPS (fp16), $5,599
The DGX Spark has 3x less memory bandwidth than the M3 Ultra but 4x more FLOPS.
By running compute-bound prefill on the DGX Spark, memory-bound decode on the M3 Ultra, and streaming the KV cache over 10GbE, we are able to get the best of both hardware with massive speedups.
Short explanation in this thread & link to full blog post below.
Running Kimi K2.5 on my desk.
Runs at 24 tok/sec with 2 x 512GB M3 Ultra Mac Studios connected with Thunderbolt 5 (RDMA) using @exolabs / MLX backend.
Yes, it can run clawdbot.