PseudoBrain-MiniMax-M3.1-Flash-285B-NVFP4

PseudoBrain (PB-27) 1M-Context Windowed-Hybrid Conversion of MiniMax-M3.1-Flash Abliterated (Blackfrost-Research/MiniMax-3.1-Flash-DeRisked)

Architecture Overview (PB27MiniMaxM3SparseForConditionalGeneration)

PseudoBrain-MiniMax-M3.1-Flash converts the 15 dense full-attention layers ([3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47, 51, 55, 59]) of MiniMax-M3.1-Flash (60 layers, 45 block-sparse lightning-indexer attention layers + 15 dense full-attention layers) into PB-27 Windowed-Hybrid Layers:

1. Windowed Attention (pb_attention_window = 16384, pb_attention_sinks = 4) + Quest/Landmark Exact Memory Tier (pb27_vllm.memory):
   • Bounds KV-cache growth on all 15 dense attention layers to a 16K sliding window while retrieving exact page-level top-$k$ KV blocks from the long-context archive up to 1,056,768+ tokens.
2. Parallel Gated DeltaNet Recurrent State Branch (self.gdn, QwenGatedDeltaNetAttention):
   • Initialized directly from the teacher's self_attn projections (pb27-gdn-00001-of-00001.safetensors) with zero-initialized output projection (out_proj = 0) so short-context KL divergence is 0.000000 while maintaining a recurrent linear-time state across 1M+ token contexts.
3. All 45 Block-Sparse Lightning-Indexer Layers, MoE Routers/Experts (NVFP4 + MXFP8), and Vision Tower Unchanged:
   • 100% lossless short-context generation and reasoning parity with the base checkpoint.

Available Editions

Repository	Base Checkpoint	Format
defnoteatin/PseudoBrain-MiniMax-M3.1-Flash-285B-NVFP4	Blackfrost-Research/MiniMax-3.1-Flash-DeRisked	Full 128-Expert Abliterated NVFP4 + MXFP8
defnoteatin/PseudoBrain-MiniMax-M3.1-Flash-NVFP4	MiniMaxAI/MiniMax-M3.1-Flash	Full 128-Expert Official NVFP4 + MXFP8
defnoteatin/PseudoBrain-MiniMax-M3.1-Flash-E32-NVFP4	Sm0r3s/MiniMax-M3.1-Flash-E32-unhealed	Compact 32-Expert (68.9 GB) NVFP4 + MXFP8
defnoteatin/PseudoBrain-MiniMax-M3.1-Flash-285B-NVFP4	primitive-ai/MiniMax-M3.1-Flash-pruned-285B-NVFP4	96-Expert (285B) NVFP4 + MXFP8

Quickstart with pb27_vllm

pip install -e ./pb27_vllm_plugin
vllm serve defnoteatin/PseudoBrain-MiniMax-M3.1-Flash-285B-NVFP4 \
  --trust-remote-code \
  --max-model-len 1056768 \
  --kv-cache-dtype fp8_e4m3