Nine of the 43 recipes on All recipes were Proven as of November 3, 2026, meaning someone other than the author reproduced them on their own cards. The fastest is GLM 5.3 Flash on four 8 GB cards at 442.1 tok/s single-user decode, and all nine run on 8 GB cards unlocked to 64 GB.
GLM 5.3 Flash on four 8 GB cards
Morrowmake's W4A16 recipe on a vLLM fork, release 1.8.0, measured 442.1 tok/s for one user on code and 3,161 tok/s cold prefill at 180 W per card, TP4 on PCIe x16 with P2P, per @voidgiraffe (October 16, 2026). The page calls it the community's reference build for four cards.
Seven confirmations back it, among them @con3po7016 (September 26, 2026; 271, 262 and 184 tok/s structured, code and prose on TP4 at Gen2 x8), @cibernox (October 13, 2026; over 350 tok/s in short bursts, 270+ on real workflows) and @turbine (October 27, 2026; 317 tok/s on code). Whether the release figures represent real use is Contested, meaning members disagree, because temperature 0 and synthetic prompts let the speculative drafter hit unusually often.
The other Proven recipes
| Recipe | Cards | Decode tok/s | Tested by | Reproduced by |
|---|---|---|---|---|
| Qwen3.8 Flash Next (ElectroHobo W4A16 TP2) | 2x 8 GB | 349.7 | @nathang6866 | @darkmaniac7 (lab notes) |
| DeepSeek V4.1 Flash (GNmac) | 4x 8 GB | 258 | @greencandle9753 | @darkmaniac7 (lab notes); partly @mountains1605 |
| Qwen3.8 27B (W4A16 vLLM) | 1x 8 GB | 169.5 | @wakeupmbrc1 | @kellertreppe123; partly @chonsworth |
| Qwen3.8 Flash Next (iIIusi0n SSD PLE) | 1x 8 GB | 110.99 | @iIIusi0n | @tekkenz, @dottoreargus, @hogg7701 |
| DeepSeek V4 Flash 0731 (allover326 PP4 DSpark) | 4x 8 GB | 98.1 | @deejay9983 | @programmerboi, @cute_sc2, @danske1 |
| Qwen3.8 Flash Next (clee PP3) | 3x 8 GB | 74.4 | @c133 | @68x0 |
| GLM 5.3 Flash (ExLlamaV3 layer split) | 2x 8 GB | 60 | @68x0 | @nestaa51 |
| GLM 5.3 (vllm-170hx-glm5) | 8x 8 GB | 48.7 | @reala7096 | @1a2b3c0967 |
Decode figures are the single-user numbers in each recipe's infobox. "Partly" marks a confirmation saved as Partly reproduced.
Reading the numbers
The Benchmarks page warns that protocols differ. Most authors use 400-token answers at temperature 0, the lab uses 1,024-token answers, and power caps, link width, P2P and the host CPU all move the result.
Some reproductions came in below the infobox. @68x0 measured 52 to 60 tok/s on the clee PP3 recipe on October 8, 2026, later about 90 tok/s, and @1a2b3c0967 confirmed the eight-card GLM 5.3 build ran on October 8, 2026 without posting tok/s.
The fastest recipe still waiting for a full reproduction is DeepSeek V4.1 Flash on 4x 8 GB (Morrowmake) at 297 tok/s; it is Reported, the badge for one person's result, with three Partly reproduced confirmations.
Sources
- All recipes
- GLM 5.3 Flash on 4x 8 GB
- Benchmarks
- Qwen3.8 Flash Next on 2x 8 GB (ElectroHobo W4A16 TP2)
- DeepSeek V4.1 Flash on 4x 8 GB (GNmac)
- Qwen3.8 27B on 1x 8 GB (W4A16 vLLM)
- Qwen3.8 Flash Next on 1x 8 GB (iIIusi0n SSD PLE)
- DeepSeek V4 Flash 0731 on 4x 8 GB (allover326 PP4 DSpark)
- Qwen3.8 Flash Next on 3x 8 GB (clee PP3)
- GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 layer split)
- GLM 5.3 on 8x 8 GB (vllm-170hx-glm5)
- DeepSeek V4.1 Flash on 4x 8 GB (Morrowmake)
- Status taxonomy