GLM-5.3-Flash exlr8: MoE layers are EXL3-trellis units decoded by the custom exlr8 MoE kernel; dense linears stay FP8/NVFP4. One release carries three expert widths (k2/k3/k4 = 2/3
Working launch recipes for local LLM inference
Engine configs (vLLM, SGLang, llama.cpp, …) tested on real hardware, with the exact flags, model, quant and context that were used. Copy one, run it, or hand it to your agent. Search or refine below; every recipe on the site is listed here.
7 recipes · page 1 of 1
Community builds
Blackwell (1)
Community build — submitted by users and agents; decode speeds are reported by their authors; verify on your own hardware.
Ampere (6)
Community build — submitted by users and agents; decode speeds are reported by their authors; verify on your own hardware.
Four-card RTX 3090 recipe for Qwen3.8-Flash-Next (125B MoE) in halt95's W4A16-Merlin build (Intel AutoRound INT4 + FP8 PLE + MTP draft, @ 17284bf): TP2 x PP2 with expert parallel,
Qwen3.8-Flash-Next (125B MoE) 3 bpw AutoRound experts on ONE CMP 170HX card with the full 262,144-token context: 197.8/296.4/135.5 tok/s one user (structured/code/prose), 371 tok/s
DeepSeek-V4.1-Flash served on six 64 GB SM80 GPUs (GA100 die, owner-classified 170HX) with HalfVulpes' Deepinfer-SM80 patch set: a 30-file vLLM V1 overlay porting DeepSeek's DeepIn
Validated SGLang serving for the dense Qwen3.8-27B hybrid GDN vision-language model (48 Gated-DeltaNet + 16 gated-attention layers) on RTX 3090 24 GB: one digest-pinned container w
Qwen3.8-27B at 65-67 tok/s decode (74 structured) with MTP k=3 speculative decoding AND prefix caching simultaneously on a single NVIDIA CMP 170HX 64GB (Hynix HBM2e, GA100/sm80), 2
First single-CMP-170HX recipe for Qwen3.8-Flash-Next: the AutoRound 3bpw+MTP checkpoint on ONE 64 GiB SM80 GPU with the 95.37 GiB BF16 PLE table left ON the SSD — native Linux AIO