PseudoBrain-Qwen3.5-9B-FP8 — 1M-Context 9B Dense Model — Runs 1,000,000 Tokens inside <6.5 GB VRAM on an 8 GB – 12 GB Consumer GPU

[!IMPORTANT] Created by Ian Kurt Galeon (@defnotean · Hugging Face @defnoteatin · iangaleonofficial@gmail.com)
GitHub & Universal pb27_vllm Engine Plugin: https://github.com/defnotean/pseudobrain
Architecture Paper: PSEUDOBRAIN_ARCHITECTURE_PAPER.pdf · PSEUDOBRAIN_ARCHITECTURE_PAPER.md · Benchmark Receipts: reports/
Open to LLM Inference Systems / Kernel & Engine Engineering roles, research collaborations, and enterprise on-premise 1M-context deployments.

---

1. Why PseudoBrain-Qwen3.5-9B Exists

In standard Transformer serving, model weights are paid once, while KV cache memory grows linearly ($O(L)$) and prefill compute grows quadratically ($O(L^2)$) — forcing long-context (128K–1,000,000 token) workloads to OOM even when the GPU easily fits the model weights.

PseudoBrain-Qwen3.5-9B-FP8 (PB27ForConditionalGeneration) solves this via the PseudoBrain architecture:

1. Hybrid Windowed-Attention + Parallel Gated DeltaNet (GDN): 32 layers: 24 native Gated DeltaNet + 8 full-attention layers ([3, 7, 11, 15, 19, 23, 27, 31]) converted into 16K SWA + Parallel Gated DeltaNet + Quest Exact Memory. Recurrent GDN state is $O(1)$ fixed size regardless of whether context is 4K or 1,000,000 tokens.
2. Three-Tier NVMe/Page-Cache Quest Exact Memory (piece_attention): On the sparse anchor layers, queries attend over [4 StreamingLLM Sinks + Super-Block Landmarks + Top-k Quest Exact-Memory Pages + 16,384 Sliding Window] via FlashAttention-2 LSE-merge (pb27_vllm/memory.py) with 2-Level Hierarchical bf16 Quest Super-Block Bounds (PB27_MEMORY_SUPERBLOCK=64).
3. Piecewise-Proportional, Strictly Non-Overlapping Block RoPE (_reposition_rope_blocks): Enforces a mandatory 64-token footprint per retrieved block plus proportional historical-gap scaling ($T(b_j) = P_{ ext{start}} + j \cdot B + ext{round}(rac{G_{ ext{cum},j}}{G_{ ext{tot}}} S_{ ext{slack}})$) — preserving exact +1 local continuity across contiguous block runs, strictly preventing phase overlap between nearby non-adjacent blocks, and maintaining 99%+ proportional macro-timeline distance fidelity at 1M tokens without static YaRN degradation.
4. Dense Global Aggregation (PB27_MEMORY_GLOBAL_AGG="auto"): Automatically detects high-participation-ratio global queries (_detect_global_agg) and activates (a) Resident Super-Block Summary Landmarks (smean_k, smean_v, 1 summary token per 4,096 tokens in GPU HBM), (b) Stratified Super-Block Timeline Coverage, and (c) Novelty-Rotating Decode Working-Set Sweeps (visit_counts) so multi-step reasoning sweeps across hundreds of distinct blocks over 1M tokens.
5. True $O(1)$ Constant-Speed Decode & Version-Resilient vLLM Plugin: Decouples the scaled prefill retrieval budget from the constant decode budget (PB27_MEMORY_DECODE_STRIDE=64, PB27_MEMORY_DECODE_BLOCKS=64), cutting 1M-token KV cache memory by 99.2% (32.0 GB → ~256 MiB active window (~1.0 GB pool)), with dual vLLM V1/V2 runner support and a zero-hook ForwardContext fallback.

Hardware & Precision Editions for PseudoBrain-Qwen3.5-9B

---

2. Two-Command Quickstart (vLLM on NVIDIA CUDA & AMD ROCm)

Option A: Install Plugin Directly into Existing vLLM Environment (2 Commands)

# 1. Install the universal PseudoBrain vLLM V1 plugin (auto-registers via vllm.general_plugins)
pip install --no-deps "git+https://github.com/defnotean/pseudobrain.git#subdirectory=serving/pb27_vllm"

# 2. Launch 1,000,000-token OpenAI-compatible server
VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve defnoteatin/PseudoBrain-Qwen3.5-9B-FP8 \
  --served-model-name pb27 \
  --max-model-len 1048576 \
  --enforce-eager \
  --no-enable-prefix-caching \
  --kv-cache-memory-bytes 6G \
  --max-num-seqs 16 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

(Note: Every Hugging Face checkpoint also bundles pb27_vllm_plugin/ directly inside the repository if you prefer pip install --no-deps -e ./pb27_vllm_plugin from a local snapshot download.)

Option B: One-Command Docker Launch (serve.sh — Auto-Detects x86_64 & aarch64)

git clone https://github.com/defnotean/pseudobrain.git
cd pseudobrain
bash serving/pb27_vllm/serve.sh defnoteatin/PseudoBrain-Qwen3.5-9B-FP8 pb27 8000

Option C: Connect OpenCode / Agentic Coding Harnesses (~/.config/opencode/opencode.jsonc)

{
  "$schema": "https://opencode.ai/config.json",
  "model": "pseudobrain/pb27",
  "small_model": "pseudobrain/pb27",
  "compaction": { "auto": true, "prune": true },
  "provider": {
    "pseudobrain": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "PseudoBrain 1M Local Server",
      "options": {
        "baseURL": "http://127.0.0.1:8000/v1",
        "apiKey": "local-loopback",
        "timeout": 600000
      },
      "models": {
        "pb27": {
          "name": "PseudoBrain-Qwen3.5-9B-FP8 (1M Context)",
          "tool_call": true,
          "limit": { "context": 1048576, "output": 16384 }
        }
      }
    }
  }
}

---

3. Verified PseudoBrain Benchmark Results (Committed Receipts in reports/)

3.1 Same-Hardware (DGX Spark GB10, TP1) Speed & Memory vs. Stock at 65K and 1,000,000 Tokens

Measured directly against upstream runtime-qualification.json on the exact same DGX Spark (GB10, 128 GB Unified Memory), TP1 hardware:

3.2 8K to 1,000,000 Token Exact Retrieval (needle@0.1/0.5/0.9, multi4, trace4)

3.3 General Intelligence, Reasoning, & Software Engineering Parity

• MMLU (n=300, 0-shot): 85.00% (255/300, within statistical stderr of stock ~85.0–87.9%).
• GSM8K (n=150, strict CoT): 96.00% (144/150, +2.71% above official 93.29% report).
• MMLU-Pro (n=150, 10-choice college reasoning): 81.33% raw / 85.31% non-truncated (vs. 73.23% Base / 80.6% Instruct).
• Algorithmic Coding (custom_code_bench, 33 medium-hard Python problems): 96.97% (32/33) on PseudoBrain-Qwen3.8-Flash-Next at 74.74 tok/s; 90.91% (60/66) on PseudoBrain-27B.
• Real-World Post-Cutoff Software Engineering (PB-SWE-Live, 59 tasks, 100-step agent loop): 26 / 59 (44.1%) vs. Stock 28 / 59 (47.5%) (paired McNemar $p \approx 0.7$, statistically indistinguishable).

---

4. Complete PseudoBrain 1M-Context Model Lineup (10 Families × 4 Precisions = 40 Models)

All 40 repositories share the single unified pb27_vllm plugin (pip install --no-deps "git+https://github.com/defnotean/pseudobrain.git#subdirectory=serving/pb27_vllm"), with slim BF16 GDN delta overlays hosted at defnoteatin/pb27-checkpoints:

---

5. License, Citation & Contact

• Code & Plugin License: Apache-2.0 (https://github.com/defnotean/pseudobrain). Model weights inherit their upstream base model license (apache-2.0).
• Author: Ian Kurt Galeon
  • GitHub: @defnotean
  • Hugging Face: @defnoteatin
  • Email (Recruiting, Research & Enterprise Inquiries): iangaleonofficial@gmail.com