MiniMax-M2.7-REAP-172B-A10B-NVFP4-GB10
MiniMax-M2.7-REAP-172B is a 172B-parameter mixture-of-experts language model, heavily quantized to ~92 GiB for deployment on NVIDIA Blackwell (GB10) GPUs. The model uses 4-bit quantization on attention layers and 8-bit on the output head, with MoE experts kept at their original precision. It is designed to run on vLLM or SGLang with minimal setup, though peak performance (≥35 tok/s decode) requires an optimized SGLang fork tuned for GB10 hardware.
Key facts
Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.
| Field | Value |
|---|---|
| Developer | scottgl |
| Parameters | 97.6B |
| Context window | Unknown |
| License | mit — OSI-approved |
| Modality / task | text-generation |
| Gated on HuggingFace | No |
| Downloads | 97.8k |
| Likes | 3 |
| Last updated | 2026-04-16 |
| Source | scottgl/MiniMax-M2.7-REAP-172B-A10B-NVFP4-GB10 |
What MiniMax-M2.7-REAP-172B-A10B-NVFP4-GB10 is
MiniMax-M2.7-REAP-172B-A10B-NVFP4-GB10 is a quantized checkpoint derived from catplusplus's M2.7 release. Architecture: 62 main layers + 3 MTP layers, 3072 hidden dim, 48 attention heads (8 KV heads, grouped query), 192 MoE experts per layer (pruned from 256 via REAP), 1536 FFN dim per expert, 200,064 vocab, 196,608 max context. Quantization: W4A4 NVFP4 (symmetric, group_size=16) on attention Q/K/V/O and all Linear layers; FP8 per-tensor on lm_head; MoE expert weights unchanged (numeric delta <1e-8 from source). Served via upstream vLLM (--quantization compressed-tensors) or SGLang; GB10-optimized SGLang fork repacks to Marlin FP4 layout at load time. Model card documents ~1–2.5% accuracy regression vs. source catplusplus (dominant term: attention F32→NVFP4).
Run MiniMax-M2.7-REAP-172B-A10B-NVFP4-GB10 locally
Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.
from transformers import pipelinepipe = pipeline("text-generation", model="scottgl/MiniMax-M2.7-REAP-172B-A10B-NVFP4-GB10")out = pipe("Explain retrieval-augmented generation in one sentence.", max_new_tokens=128)print(out[0]["generated_text"])Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.
How you'd run it
A typical self-hosted path — open weights, an inference server, your application.
DEV.co builds each layer — from GPU infrastructure to the application.
Best use cases
Running & fine-tuning it
Estimated ~92 GiB model weights (19 shards, ~5 GiB each). On a single GPU: minimum 96 GiB VRAM (H100, L40S, or GB10 with headroom for KV cache and batch buffers). Multi-GPU: distribute via tensor parallelism. KV cache overhead depends on batch size and sequence length; for batch_size=1, 196K context ≈ 12–16 GiB additional. SGLang fork on GB10 achieves ≥35 tok/s decode; upstream vLLM/SGLang on older GPUs will be slower but functional.
Model is fully quantized (W4A4 and FP8); fine-tuning requires dequantization and retraining. LoRA/QLoRA are not mentioned in the card and are not recommended on 4-bit activations. If fine-tuning is required, use the original catplusplus checkpoint or a less-quantized variant. The card provides a conversion script (convert.py) to understand the quantization process but does not provide a dequantization pathway.
When to avoid it — and what to weigh
- Strict accuracy requirements without re-evaluation — Model card projects ~1–3% regression on benchmarks due to attention quantization. Code and reasoning tasks may see ~3% loss. Recommend benchmark validation before production use.
- CPU or non-Blackwell GPU deployment — Optimized for GB10/Blackwell (SM 12.1). While upstream vLLM/SGLang can load it, peak throughput requires Blackwell; older GPUs (A100, H100) will not auto-repack weights and may see degraded performance.
- Fine-tuning or continued pretraining — Quantized model; not suitable for LoRA/QLoRA without dequantization and re-training infrastructure. Originating checkpoint (catplusplus) should be used if fine-tuning is needed.
- Real-time inference with sub-second latency SLA — MoE expert selection and dynamic routing add latency. Model is optimized for throughput (tok/s), not latency. Verify against your SLA before adoption.
License & commercial use
MIT license (inherited from upstream MiniMax-M2). MIT is a permissive OSI-approved open-source license allowing commercial use, modification, and distribution with minimal restrictions.
MIT license permits unrestricted commercial use provided the license and copyright notice are retained. No gating or restricted access. However, verify that the original MiniMax-M2 base model (and REAP pruning provenance) comply with your organization's IP policies. Attribution to catplusplus and MJPansa (REAP methodology) is requested in documentation.
DEV.co evaluation signals
Editorial assessment — not user reviews. Directional, with an explicit confidence level.
| Signal | Assessment |
|---|---|
| Maintenance | Active |
| Documentation | Adequate |
| License clarity | Clear |
| Deployment complexity | Moderate |
| DEV.co fit | Strong |
| Assessment confidence | High |
Model uses compressed-tensors quantization format (version 0.14.1.dev0); ensure hosting platform trusts remote code (--trust-remote-code flag required). No explicit security audit or adversarial robustness claims. Quantization introduces model compression; verify that the 1–3% accuracy regression does not affect safety-critical use cases (e.g., autonomous systems). No data provenance or filtering claims beyond base MiniMax-M2. Standard LLM considerations apply: prompt injection, hallucination, and output filtering are user's responsibility.
Alternatives to consider
catplusplus/MiniMax-M2.7-REAP-172B-A10B-NVFP4 (source checkpoint)
Original unquantized-on-attention model; better accuracy (~1–3% higher) but larger (~101 GiB). Use if accuracy is critical and hardware supports it.
saricles/MiniMax-M2.5-REAP-172B-A10B-NVFP4-GB10
Similar quantization and GB10 optimization but 2.5B parameters vs. 2.7B. Smaller, faster inference; slightly lower capability. Use if latency is primary constraint.
Meta Llama (various quantized variants)
Broader ecosystem support, more diverse serving tools (TGI, llama.cpp, Ollama). Less specialized for Blackwell but more portable across hardware.
Ship MiniMax-M2.7-REAP-172B-A10B-NVFP4-GB10 with senior software developers
Get started with high-throughput inference on NVIDIA Blackwell using our curated SGLang setup guide. Benchmark the model on your workload, review the 1–3% accuracy regression estimates, and integrate into your inference stack with MIT-licensed commercial freedom.
Talk to DEV.coRelated open-source tools
Surfaced by semantic similarity across the DEV.co open-source index.
Related on DEV.co
Explore the category and the services that help you build with it.
MiniMax-M2.7-REAP-172B-A10B-NVFP4-GB10 FAQ
Can I use this model for commercial applications?
What GPU do I need to run this model?
How much accuracy is lost compared to the original catplusplus model?
Can I fine-tune this quantized model?
Work with a software development agency
Adopting MiniMax-M2.7-REAP-172B-A10B-NVFP4-GB10 is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.
Deploy Quantized MiniMax-M2.7 on Blackwell Today
Get started with high-throughput inference on NVIDIA Blackwell using our curated SGLang setup guide. Benchmark the model on your workload, review the 1–3% accuracy regression estimates, and integrate into your inference stack with MIT-licensed commercial freedom.