NVFP4 quant makes MiniMax M2.5 REAP practical
This is a community-uploaded NVFP4 quantization of Cerebras’s REAP-pruned MiniMax-M2.5 checkpoint, tuned for NVIDIA DGX Spark / GB10-class 128GB Blackwell systems. The upload claims the model runs on a single device with no source patches, using compressed-tensors plus a specific vLLM setup, and it keeps the core appeal of MiniMax-M2.5: a 172B sparse MoE model with 10B active parameters, long context, and strong coding/tool-use capability. The post is essentially a deployment unlock for people trying to run a high-end agentic model locally on Blackwell hardware.
Hot take: this is more interesting as a hardware enablement release than as a new model release. It takes an already strong open-weight model and packages it into something that appears usable on a very specific class of 128GB Blackwell machines.
- –Base model lineage is solid: MiniMax-M2.5 -> Cerebras REAP 172B -> NVFP4 community quant.
- –The pitch is pragmatic, not theoretical: single-box deployment, vLLM-compatible, and benchmarked on DGX Spark GB10.
- –The main caveat is portability: the author explicitly notes YMMV, especially on NVIDIA Thor, so this is not a universal “it just works” release.
- –The post suggests good real-world competence, but the evidence is still anecdotal from a coding test rather than broad evals.
- –For local LLM users, the value is in making a frontier-ish MoE model accessible on high-memory consumer/prosumer Blackwell systems.
DISCOVERED
53d ago
2026-04-04
PUBLISHED
53d ago
2026-04-04
RELEVANCE
AUTHOR
catplusplusok