Back to README.md
Clover-35B-A3B-Preview by DiffuSpace, a file of 13 to 21 GB. It is Qwen3.6-35B-A3B trained further to also write a block of tokens at once. It reads text only.
Measured for PolyStrata on the laptop, with a short chat and a prompt of 32,000 tokens. Its card has 8 GB, and the processor runs the experts the card has no room for.
| NVIDIA: RTX 4070 Laptop (8 GB), Core i7-14700HX, 32 GB RAM | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
After the prompt of 32,000 tokens the laptop writes at 60 to 65 tokens/s with Q3_K_M. Q4_K_M fills the 32 GB of RAM there, which is what makes it slow and uneven. The workstation with the RTX 5090 was not measured.
Qwen3.6-35B-A3B writes faster on the same laptop, 94 to 100 tokens/s, because its file has a draft block that guesses the next words for one layer's work. Clover's files have none. Clover can guess a block of words itself, but that runs the whole model once more, and PolyStrata leaves it off: on a card that held every expert the answers came sooner without it (measured).
On an RTX PRO 6000 that held every expert of Q3_K_M, PolyStrata picked the same next token as llama.cpp (the
clover_moe support of its pull request 31306) at 127 of 128 steps after a short prompt and at 125 of 128 after
5,573 tokens.
| Graphics card | One NVIDIA GeForce RTX 20 series card or newer with 8 GB of VRAM or more. About 3.5 GB are taken before any expert. It has run on a laptop's 8 GB card and on a 96 GB card. It has not been tried on an AMD card. |
| RAM | About 16, 20 or 25 GB beside an 8 GB card for APEX-I-NANOPLUS, Q3_K_M and Q4_K_M, the installer's estimate. A card with 24 GB holds the two smaller sizes whole. All three have run on a laptop with 32 GB. |
| Disk | 13, 17 or 21 GB for the model file. |
| System | Linux or Windows 10 / 11, and a current NVIDIA driver. |
The installer sets up everything else, as for Qwen3.6-35B-A3B. maternion's files carry no chat template, so the installer downloads the model's own from DiffuSpace. Clover's license is Apache 2.0.
Three sizes, all made with the converter of llama.cpp's pull request 31306. The installer takes Q3_K_M, or
APEX-I-NANOPLUS when the PC has too little RAM for it. Q4_K_M is chosen by name: --model Q4_K_M.
| Size | What it is | Download | RAM beside an 8 GB card |
|---|---|---|---|
| Q3_K_M | the whole model with 3-bit experts (maternion) | 17 GB | about 20 GB |
| APEX-I-NANOPLUS | 2- and 3-bit experts from an importance matrix (IsValorum): the least RAM | 13 GB | about 16 GB |
| Q4_K_M | 4-bit experts (maternion) | 21 GB | about 25 GB |
maternion's two sizes were quantized without an importance matrix. How close each size comes to the model as trained was not measured.
The diagram and how it works: HOW_IT_WORKS_MODELS.md.