Two-Spark MiniMax Files Arrive Before Their Numbers
A 3.06 bpw EXL3 pack and a 203.5 GB GGUF fit 256 GB; neither has a Spark result yet.
Downloads first, measurements later.
@MiaAI_lab posted that a MiniMax-M3.1-Flash TensorFold recipe for two DGX Sparks is “dropping soon,” with her EXL3-EAQ quant ready. Its card lists 3.064 bits per weight with every routed expert at 3 bits, the sparse-attention indexer at FP16 and the router at FP32. The text model is about 167 GB; the repo, with draft and vision tower, is about 171 GB.
The card is plain about what is missing. No quality number versus the original. No engine verified to load it. No long-context or speculative-decoding test.
AtomicChat released GGUFs of the 428B model “from 258GB down to 203.5GB” and recommends its 203.5GB Q3_K_M for 256GB hardware.
| File | On disk | Model memory | Fits |
|---|---|---|---|
| Q2_K | 155.9 GB | 145.2 GiB | 192 GB machines |
| Q3_K_M | 203.5 GB | 189.5 GiB | 256 GB machines |
| Q4_K_M | 258.0 GB | 240.3 GiB | 384 GB machines and up |
@ViC305 says he is working on a one-Spark EXL3 quant with his SAGE-EXL3 process. In a later reply, Mia wrote: “Actually I think I’m gonna skip the recipe and focus on TF with 4.” By our 6:51 AM PT gather, no two-Spark recipe had appeared.
Sources
- @MiaAI_lab on X · Oct 22, 11:58 PM PT
- @MiaAI_lab on X · Oct 23, 1:08 AM PT
- @atomic_chat_hq on X · Oct 23, 4:58 AM PT
- @ViC305 on X · Oct 22, 9:44 PM PT
- Mia-AiLab MiniMax EXL3 3.06 bpw card · huggingface.co
- AtomicChat MiniMax GGUF card · huggingface.co