Recipe finder
From CMP Wiki
More actions
Find a recipe
Choose a quick search below, or use the filter button to select your cards, model and links. All recipes is the complete sortable index; the Setup matrix starts from a hardware configuration. Recipes keeps model-selection notes and history.
Find a recipe that fits your cards. Choose the card variant, how many cards you have, the model, the engine and what your links can do (P2P, PCIe generation and width); an empty box matches anything. The results list every matching recipe page with its numbers and keep decode with one user, total decode with many users and prefill apart, because they measure different things. "not stated" means the recipe page does not give that value: open the recipe, and add the value there if you know it.
Have a whole setup in mind (card count, memory per card, PCIe link, P2P, ECC)? The Setup matrix sorts every recipe by whether it runs on that setup and labels each speed as measured, estimated or unknown.
Quick searches
How to read the results
- 1 user tok/s: decode speed with one request at a time. Where the author posted three numbers (structured, code, prose), this is the code one. Hover a number for how it was measured and which speculative decoding was on.
- Many users tok/s: decode summed over all concurrent requests, with the number of requests in brackets. It shows how much a setup can serve, not how fast one chat feels.
- Prefill tok/s: prompt reading speed, with the prompt length in brackets where the recipe states it.
- P2P: needed (the recipe needs GPU peer-to-peer), optional (it works with or without), on (it was run with P2P on; whether it needs it is not stated), not used (it runs without P2P). Single-card recipes count as not used.
- PCIe: the link generation and width the numbers were measured on. Stock cards run at x4; x16 needs the PCIe x16 capacitor mod.
- Numbers from different protocols, power caps and links are not directly comparable, and most recipes use speculative decoding, which flatters benchmark prompts against real work. Benchmarks has every measurement with its setup.
- Status is computed: Proven means someone else reproduced the recipe (see Status taxonomy).
Every recipe
The full table, sorted by decode with one user. Click a column heading to sort by another column.
47 recipes in the table. Sorted by decode with one user; click a column heading to sort again. "not stated" means the recipe page does not give that value.
On a phone this table shows the recipe, the three speeds and the status. Turn the phone sideways or open a recipe for the rest.
| Recipe | Model | Quant | Engine | Cards | Topology | Context | 1 user tok/s | Many users tok/s | Prefill tok/s | P2P | PCIe | Power | Status | Tested on |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GLM 5.3 Flash on 4x 8 GB | GLM 5.3 Flash | W4A16 (INT4 weights, group size 128, FP16 activations): canada-quant/GLM-5.3-Flash-W4A16-MTP | vLLM | 4x 8 GB | TP4 (default) or PP4 | 262,144 | 442.1 | 835 (8) | 3161 | optional | Gen2 x16 | 180 W | Proven | 2026-10-16 |
| Qwen3.8 Flash Next on 2x 8 GB (ElectroHobo W4A16 TP2) | Qwen3.8 Flash Next | W4A16: halt95/Qwen3.8-Flash-Next-W4A16-Merlin (AutoRound INT4 g128 experts, INT8 GDN projections, FP8 n-gram table) | vLLM | 2x 8 GB | TP2 with whole-expert expert parallel | 262,144 | 349.7 | 1937 (128) | 4696 | needed | Gen2 x16 | 180 W | Proven | 2026-10-23 |
| DeepSeek V4.1 Flash on 4x 8 GB (Morrowmake) | DeepSeek V4.1 Flash | Hybrid: routed experts EXL3 2.9 bpw (LibertAIDAI), everything else official FP8/BF16; KV BF16 | vLLM | 4x 8 GB | TP4; 1.0.1 splits every routed expert across the four cards (1.0.0: 96 whole experts per card) | 262,144 | 297 | 795 (8) | 3776 (2,048 ) | optional | x16 | 180 W | Reported | 2026-11-01 |
| Qwen3.8 Flash Next on 1x 8 GB (ElectroHobo 3 bpw) | Qwen3.8 Flash Next | AutoRound mixed 2/3-bit (klee100, 2.998 effective bpw, 47 GiB on card) plus halt95's FP8 n-gram table in host RAM | vLLM | 1x 8 GB | single card | 262,144 | 296.4 | 1384 (48) | 4243 | not used | Gen2 x16 | 180 W | Reported | 2026-10-23 |
| DeepSeek V4.1 Flash on 4x 8 GB (GNmac with copy fix and faster prefill) | DeepSeek V4.1 Flash | Hybrid as recipe 1.2.0: routed experts EXL3 from LibertAI's 2.9 bpw pack (mostly 3 bpw, 2 bpw on layers 18-22); attention, shared experts, indexer and DSpark drafter official FP8; KV fp8_ds_mla | vLLM | 4x 8 GB | TP4 (LAYOUT=tp4); the prefill changes are TP4 only | 262,144 | 283 | 559 (8) | 2732 (24k, 34k and 38k , cold, median of 9) | needed | Gen2 x16 | 180 W | Reported | 2026-11-09 |
| MiMo V2.6 Flash on 4x 8 GB | MiMo V2.6 Flash | NVFP4 (ModelOpt mixed precision): ProCreations/MiMo-V2.6-Flash-RL-NVFP4, run as 4-bit weights with BF16 activations through Marlin | vLLM | 4x 8 GB | TP4 (P2P on); expert parallel off | 1,048,576 | 274.9 | 738 (8) | 2722 (24K to 38K ) | needed | Gen2 x16 | 180 W | Reported | 2026-10-18 |
| DeepSeek V4.1 Flash on 4x 8 GB (GNmac) | DeepSeek V4.1 Flash | Hybrid: routed experts EXL3 (mostly 3 bpw, 2 bpw on layers 18-22) from LibertAI's 2.9 bpw pack; attention, shared experts, indexer and DSpark drafter official FP8; KV fp8_ds_mla | vLLM | 4x 8 GB | TP4 (default) or PP4 for x1/x4 links | 262,144 | 258 | 544 (8) | 2416 | needed | Gen2 x16 | 180 W | Proven | 2026-10-26 |
| Qwen3.8 Flash Next on 2x 8 GB (mannydelrio PP2) | Qwen3.8 Flash Next | AWQ W4A16 g32 (leoncca/Qwen3.8-Flash-Next-Uncensored-AWQ-g32); PLE table converted to INT4 and served from SSD | vLLM | 2x 8 GB | PP2 (layer partition 25,23), no P2P | 262,144 | 177.8 | 734.2 (8) | 7006 (24K to 38K ) | not used | Gen2 x4 | 180 W | Reported | 2026-10-15 |
| Qwen3.8 Flash Next on 2x 8 GB (Qwen-Flash-SM80, GDS TEP2) | Qwen3.8 Flash Next | NVFP4 main weights through Marlin (no native FP4 on sm_80), mixed FP8 and BF16 MTP weights, FP8 PLE table (about 51.2 GB) kept on NVMe | vLLM | 2x 8 GB | TEP2 (TP2 plus expert parallel), MTP 6; one card at PCIe Gen2 x16 (x16 mod), one at Gen2 x4; GPU P2P required | 262,144 | 168.29 | 749.67 (10) | 1986.38 (8K ) | needed | Gen2 mixed | not stated | Reported | 2026-09-08 |
| DeepSeek V4.1 Flash on 6x 8 GB (Deepinfer-SM80 TP2 x PP3) | DeepSeek V4.1 Flash | Official checkpoint (FP8 dense, FP4 experts); software FP8 on sm_80 | vLLM | 6x 8 GB | TP2 x PP3, partition 14,14,12 | 1,048,576 | 160.8 | not stated | 4037 (16K ) | not stated | not stated | not stated | Reported | 2026-10-28 |
| DeepSeek V4.1 Flash heretic on 4x 8 GB | DeepSeek V4.1 Flash | Hybrid: routed experts EXL3 2.9 bpw (LibertAIDAI), everything else official FP8/BF16 with the 38 heretic attn.wo_b tensors | vLLM | 4x 8 GB | TP4 | 262,144 | 158 | 323 (8) | 2453 (17k ) | not stated | not stated | not stated | Reported | 2026-10-29 |
| DeepSeek V4.1 Flash on 8x 8 GB (Deepinfer-SM80 TP4 x PP2) | DeepSeek V4.1 Flash | Official checkpoint (FP8 dense, FP4 experts); software FP8 on sm_80 | vLLM | 8x 8 GB | TP4 x PP2, partition 20,20 | 1,048,576 | 153.35 | 620.22 (12) | 3463 | not stated | not stated | not stated | Reported | 2026-10-28 |
| Qwen3.8 Flash Next on 2x 8 GB (Morrowmake EXL3) | Qwen3.8 Flash Next | EXL3 5.05 bpw; PLE (n-gram) table read from SSD | vLLM | 2x 8 GB | TP2 with expert parallel | 262,144 | 152.6 | 674.7 (8) | 2651 | needed | x16 | not stated | Reported | 2026-10-16 |
| DeepSeek V4.1 Flash on 4x 8 GB (CurtisAccelerate TC-MoE) | DeepSeek V4.1 Flash | Hybrid: LibertAI EXL3 2.9 bpw routed experts (layers 0-39), official FP8/FP4 for attention, shared experts, head, draft and Engram; KV fp8_ds_mla | vLLM | 4x 8 GB | TP4, peer access off (NCCL_P2P_DISABLE=1), host all-reduce | 262,144 | 144.2 | 271.6 (8) | 1523 (2,048 ) | not used | Gen2 x16 | 140 W | Reported | 2026-10-22 |
| Qwen3.8 Flash Next on 2x 8 GB (DUFRelic PP2) | Qwen3.8 Flash Next | W4A16 INT4 g128 with the PLE table repacked INT4 g32: zankich/Qwen3.8-Flash-Next-W4A16-Merlin-INT4PLE-FP8KV; KV BF16 | vLLM | 2x 8 GB | PP2, MTP depth 3 | 262,144 | 133.39 | 245.17 (4) | 6213.66 (16384 ) | optional | not stated | 140 W | Reported | 2026-10-03 |
| GLM 5.3 Flash on 8x 8 GB (promisezackr PP8 NVFP4 DFlash2) | GLM 5.3 Flash | LibertAIDAI/GLM-5.3-Flash-NVFP4 (ModelOpt NVFP4, routed experts only, 181 GiB), run through Marlin W4A16 kernels; GLM-5.3-Flash DFlash2 drafter (5 sliding-window layers) | vLLM | 8x 8 GB | PP8 (VLLM_PP_LAYER_PARTITION=7,5,7,5,7,5,7,2) with DFlash2, 7 draft tokens; PCIe Gen2 x4, no P2P; host 2x Xeon E5-2690 v4, 96 GB DDR4 | 1,048,576 | 123 | 276 (16) | not stated | not used | Gen2 x4 | not stated | Reported | 2026-08-31 |
| DeepSeek V4.1 Flash on 8x 8 GB (Zanooda PP8) | DeepSeek V4.1 Flash | Official checkpoint as downloaded: MXFP4 routed experts as Marlin W4A16, MXFP8 dense as Marlin W8A16; KV fp8 | vLLM | 8x 8 GB | PP8, partition 5,5,5,5,5,5,5,5 | 1,048,576 | 117 | 560 (8) | 6066 (104,881 ) | optional | Gen2 mixed | 180 W (one card unstable at 250 W) | Reported | 2026-10-19 |
| Qwen3.8 Flash Next on 4x 8 GB | Qwen3.8 Flash Next | AWQ W4A16: wtdcode/Qwen3.8-Flash-Next-AWQ-W4A16 | vLLM | 4x 8 GB | TP4 with expert parallel, MTP 3 | 1,000,000 | 116.49 | not stated | 2101.4 (42,584 ) | not stated | not stated | not stated | Reported | 2026-10-16 |
| Qwen3.8 Flash Next on 1x 8 GB (iIIusi0n SSD PLE) | Qwen3.8 Flash Next | AutoRound 3 bpw + MTP: klee100/Qwen3.8-Flash-Next-AutoRound-3bpw-MTP; 95.37 GiB BF16 PLE table on SSD | vLLM | 1x 8 GB | single card | 262,144 | 110.99 | 667 (16) | 2636 (8K ) | not used | not stated | 180 W | Proven | 2026-09-19 |
| GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 TP2) | GLM 5.3 Flash | EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3; DFlash2 drafter converted to EXL3 6 bpw | ExLlamaV3 | 2x 8 GB | TP2 over NCCL (GPU P2P), DFlash2 K7 plus hybrid drafting | 392,960 | 107.67 | 67.25 (2) | 1533 (32K ) | optional | Gen2 x16 | 180 W (author's cards); about 170 W with a modest memory OC in his Discord numbers | Reported | 2026-10-13 |
| DeepSeek V4.1 Flash on 4x 8 GB (alanoo fidelity build) | DeepSeek V4.1 Flash | Hybrid: all 40 expert layers at 3 bits (LibertAI 2.9 bpw pack with layers 18-22 grafted from the 3.5 bpw pack), official FP8 dense tensors | vLLM | 4x 8 GB | TP4, P2P on | 262,144 | 105 | 482.4 (8) | 2514 (24k to 38k ) | needed | Gen2 x16 | 200 W | Reported | 2026-10-27 |
| GLM 5.3 on 8x 8 GB plus 2x RTX PRO 6000 (vLLM TP2 x PP5) | GLM 5.3 | cebeuq's hybrid: the published NVFP4 routed experts unchanged plus an INT8-g128 residue, converted from dealignai/GLM-5.3-ABLITERATED-NVFP4; about 420 GB on disk | vLLM | 8x 8 GB | TP2 x PP5 over two hosts: stage 0 = two RTX PRO 6000 Blackwell (layers 0-29 plus embedding), stages 1-4 = four CMP 170HX pairs (12 layers each); direct 100 GbE between the hosts | 524,288 | 104 | 244 (8) | 3400 (17k , cold, no prefix cache) | on | Gen2 x16 | not stated | Reported | 2026-11-05 |
| DeepSeek V4.1 Flash on 4x 8 GB plus RTX PRO 6000 (ExLlamaV3 TP5) | DeepSeek V4.1 Flash | EXL3 3.5 bpw (LibertAIDAI; routed experts 4 bpw on layers 0-9 and 30-39, 3 bpw on 10-29; attention and shared experts 6 bpw) plus the official Engram shards | ExLlamaV3 | 4x 8 GB | TP5 across four CMP 170HX and one RTX PRO 6000 Blackwell (native backend, reduces through host memory); MTP draft on the PRO | 1,048,576 | 99.9 | not stated | 2116 (26,296 ) | not used | x16 | 150 W | Reported | 2026-10-20 |
| DeepSeek V4 Flash 0731 on 4x 8 GB (allover326 PP4 DSpark) | DeepSeek V4 Flash | Native MXFP4 experts + FP8 attention (official deepseek-ai checkpoint, about 155.4 GiB); KV cache fp8 | vLLM | 4x 8 GB | PP4 (VLLM_PP_LAYER_PARTITION=12,12,12,7) with DSpark, 5 draft tokens; PCIe Gen2 x4, no P2P | 1,047,736 | 98.1 | 712.8 (64) | 5300 (25k to 77k ) | not used | Gen2 x4 | 180 W | Proven | 2026-08-05 |
| Qwen3.6 27B on 1x 10 GB (SGLang W8A8) | Qwen3.6 27B | INT8 W8A8 (compressed-tensors, dynamic) | SGLang | 1x 10 GB | single card, in a KVM virtual machine | 24,576 | 90.26 | 269.31 (4) | 4122.16 (2048 ) | not used | Gen1 x4 | 250 W | Reported | 2026-07-31 |
| Qwen3.8 Flash Next on 3x or 4x 8 GB (zebgop-ops FP8 PP) | Qwen3.8 Flash Next | Official block FP8, about 126 GiB including the about 50 GiB FP8 n-gram (PLE) table, which is served from host RAM; KV cache bf16 (the QSA layers refuse fp8) | vLLM | 4x 8 GB | PP4, VLLM_PP_LAYER_PARTITION=12,12,12,12, MTP 3, FULL_AND_PIECEWISE CUDA graphs; PCIe Gen2 x4 per card, no P2P. Three cards: PP3 with 16,17,15 or 16,16,16 | 262,144 | 89 | 442 (8) | 11100 (32k ) | not used | Gen2 x4 | 200 W | Reported | 2026-09-02 |
| DeepSeek V4.1 Flash on 6x 8 GB (TP6) | DeepSeek V4.1 Flash | Official checkpoint: dense MXFP8 through an SM80 Marlin W8A16 path, routed experts MXFP4 through W4A16; KV FP8 | vLLM | 6x 8 GB | TP6 (virtual-head padding), PP1 | 1,048,576 | 78.55 | not stated | not stated | not stated | Gen2 x16 | not stated | Reported | 2026-10-19 |
| Qwen3.8 Flash Next on 2x 8 GB (vektorprime PP2) | Qwen3.8 Flash Next | AWQ INT4: cyankiwi/Qwen3.8-Flash-Next-AWQ-INT4; 95 GiB BF16 PLE table in host RAM (FP8 table optional) | vLLM | 2x 8 GB | PP2 with PLE CPU offload | 262,144 | 75.4 | not stated | not stated | not stated | not stated | not stated | Reported | 2026-08-29 |
| Qwen3.8 Flash Next on 3x 8 GB (clee PP3) | Qwen3.8 Flash Next | W4A16 AutoRound: Intel/Qwen3.8-Flash-Next-W4A16-AutoRound; 95.4 GiB BF16 PLE table streamed from NVMe; KV fp8 e4m3 (software-decoded on SM80) | vLLM | 3x 8 GB | PP3, no P2P (works at PP2 on two cards) | 1,048,576 | 74.4 | not stated | 4831 (64,845 ) | not used | x4 | not stated | Proven | 2026-10-10 |
| GLM 5.3 on 8x 8 GB (cebeuq glm53-cmp170hx-recipe) | GLM 5.3 | hybrid: the published NVFP4 routed experts unchanged plus an INT8-g128 residue, converted offline from dealignai/GLM-5.3-ABLITERATED-NVFP4 (432.9 GiB source) | vLLM | 8x 8 GB | TP2 x PP4, layers 22,20,20,16, the MTP layer on the last stage; each TP2 pair behind one PCIe switch | 524,288 | 68.3 | 135.5 (8) | not stated | needed | Gen2 x16 | 160 W | Reported | 2026-10-10 |
| GLM 5.3 on 8x 8 GB (vllm-170hx-glm5) | GLM 5.3 | GLM-5.3-Int4-Int8Mix-AWQ-g64; drafter incoai/GLM-5.3-DFlash2 | vLLM | 8x 8 GB | PP8 (also PP9 and PP10), TP1 | 1,048,576 | 48.7 | 174.1 (8) | 1833 (about 7.8K ) | not stated | Gen2 x4 | 170 W | Proven | 2026-09-29 |
| GLM 5.2 on 8x 8 GB | GLM 5.2 | 4-bit, posted as int4 and q4 (the checkpoint is named in Notes) | vLLM | 8x 8 GB | PP8 (pipeline parallel); tensor parallel ruled out at Gen1 x4 | not stated | 30 | not stated | 2600 | not stated | Gen1 x4 | not stated | Reported | 2026-07-27 |
| GLM 5.3 on 8x 8 GB (Int4-Int8Mix PP8) | GLM 5.3 | GLM-5.3-Int4-Int8Mix (INT4 / INT8 mix); KV fp8_ds_mla | vLLM | 8x 8 GB | PP8, TP1 | 1,048,576 | 25 | not stated | not stated | not stated | not stated | 200 W | Reported | 2026-10-07 |
| Mistral Large 4 on 8x 8 GB (llama.cpp lab trial) | Mistral Large 4 | unsloth UD-IQ3_XXS GGUF (394 GiB in 10 files); DSpark drafter Q8_0 (2.1 GB) | llama.cpp | 8x 8 GB | Layer split over all eight cards (tensor split is not available for this architecture) | 196,608 | 18.7 | 34.1 (8) | 131 (17k ) | on | not stated | 250 to 300 W (governor), +160 MHz offset | Reported | 2026-11-01 |
| Qwen3.8 Flash Next on 2x 8 GB (Plean PP2) | Qwen3.8 Flash Next | Mixed INT4 AutoRound (Minachist/Qwen3.8-Flash-Next-INT4-Mixed-AutoRound, about 65 GB) with the full BF16 PLE table on NVMe | vLLM | 2x 8 GB | PP2, MTP 3, no P2P | 262,144 | not stated | not stated | 5800 | not used | not stated | not stated | Proven | 2026-10-07 |
| Qwen3.8 27B on 1x 8 GB (W4A16 vLLM) | Qwen3.8 27B | W4A16 (INT4 weights, Marlin kernels) | vLLM | 1x 8 GB | single card | 262,144 | not stated | not stated | not stated | not used | not stated | not stated | Proven | 2026-08-30 |
| GLM 5.3 Flash on 4x 8 GB (kaka86mm EXL3 + Marlin sidecar + DFlash2) | GLM 5.3 Flash | EXL3 4 bpw (Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw, 164 GB) plus an offline-generated Marlin INT4 decode sidecar (151 GB) | vLLM | 4x 8 GB | PP4 (VLLM_PP_LAYER_PARTITION=13,11,11,10), DFlash2 k=7; PCIe Gen2 x4, no P2P | 524,288 | not stated | 199 (4) | 2077 (131k ) | not used | Gen2 x4 | 200 W | Reported | 2026-09-19 |
| DeepSeek V4 Flash 0731 on 3x 8 GB (PP3 DSpark) | DeepSeek V4 Flash | Native MXFP4 experts + FP8 attention (official deepseek-ai checkpoint, about 155.4 GiB); KV cache fp8 | vLLM | 3x 8 GB | PP3 (VLLM_PP_LAYER_PARTITION=15,15,13) with DSpark, 5 draft tokens; PCIe Gen2 x4, no P2P | 691,968 | not stated | not stated | not stated | not used | Gen2 x4 | 100 W | Reported | 2026-08-15 |
| GLM 5.3 Flash on 2x 8 GB (Project Maya) | GLM 5.3 Flash | Maya-M GGUF (116 GB), every expert in VRAM; Maya-S (96.5 GB) and Maya-S24 (94.7 GB) also reported | Project Maya | 2x 8 GB | engine-managed split with expert tiering across VRAM, RAM and NVMe | not stated | not stated | not stated | 514 (8K ) | not stated | Gen2 x4 | not stated | Reported | |
| GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 layer split) | GLM 5.3 Flash | EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3 | ExLlamaV3 | 2x 8 GB | layer split (no tensor parallel), no P2P | 392,960 | not stated | not stated | not stated | not used | not stated | not stated | Proven | 2026-10-20 |
| Laguna-S-2.1 on 2x 10 GB | Laguna-S-2.1 | NVFP4 GGUF (runs through a fallback path: the cards have no native FP4) | llama.cpp | 2x 10 GB | tensor split 1,1 across the two cards | 131,072 | not stated | not stated | not stated | not stated | not stated | not stated | Reported | 2026-08-12 |
| DeepSeek V4 Flash pruned on 4x 10 GB | DeepSeek V4 Flash | vLLM | 4x 10 GB | not tensor parallel (the split is not stated) | not stated | not stated | not stated | not stated | not stated | Gen2 x4 | 250 W limit (read on 2026-08-11, during his llama.cpp runs) | Reported | 2026-08-12 | |
| DeepSeek V4.1 Flash on 2x 8 GB (Project Maya) | DeepSeek V4.1 Flash | GSQ-RCO 2-bit GGUF (this row); Maya GGUF (Maya-S 340 GB or Maya-XS 277 GB; Engram tables stay on SSD) | Project Maya | 2x 8 GB | engine-managed expert tiering across VRAM, host RAM and NVMe | not stated | not stated | not stated | 1190 (4K ) | not stated | not stated | not stated | Reported | 2026-10-28 |
| Qwen3.6 27B on 1x 10 GB (llama.cpp MTP) | Qwen3.6 27B | UD-Q4_K_XL GGUF (UD-Q8_K_XL also tested) | llama.cpp | 1x 10 GB | single card, whole model in VRAM | not stated | not stated | not stated | 496.85 (2048 ) | not used | Gen1 x4 | 250 W | Reported | 2026-07-22 |
| GLM 5.2 on 3x 10 GB (llama.cpp CPU offload) | GLM 5.2 | Q4 GGUF (the exact Q4 file is not named); UD-Q2_K_XL also tested | llama.cpp | 3x 10 GB | partial offload: part of the model on the three cards, the rest in host RAM | not stated | not stated | not stated | 48.22 (2048 ) | not stated | Gen2 x4 | not stated | Reported | 2026-07-24 |
| DeepSeek V4.1 Flash on 5x 8 GB (PP5 full residency) | DeepSeek V4.1 Flash | Official checkpoint (510 GB: MXFP4 routed experts, FP8 dense layers, 203 GB FP8 Engram table); KV cache fp8_ds_mla | vLLM | 5x 8 GB | PP5 with an even 8,8,8,8,8 layer split; the repo's cards are on PCIe Gen2 x16 with no NVLink | 1,048,576 | not stated | not stated | 4600 (60K ) | not stated | Gen2 x16 | not stated | Reported | 2026-09-22 |
| GLM 5.3 Flash on 3x 8 GB | GLM 5.3 Flash | W4A16, about 4.3 bpw, 163 GiB on disk | vLLM | 3x 8 GB | PP3 | about 1,000,000 (8-bit KV cache) | not stated | not stated | not stated | not stated | not stated | not stated | Reported | 2026-10-19 |
Browse everything
All recipes (every recipe in one sortable table) · Benchmarks (every measurement, including runs without a recipe page) · Drilldown (filter on any field) · Recipes (overview and quick picks).
Missing a value? Edit the recipe's infobox (the finder fields are described on Template:Recipe) or open the recipe in Form:Recipe; the finder shows the change right away.