README.md

---
license: apache-2.0
language:
  - en
library_name: transformers
pipeline_tag: text-generation
base_model: Cloudflare/clef-flash
datasets:
  - TypeScript/FastBrain-Instruct
tags:
  - fastbrain
  - clef
  - gguf
  - llama-cpp
  - systemone
  - decision-index
  - tool-calling
  - function-calling
  - intent-classification
  - structured-decision
---

<div align="center">

# FastBrain-9B

**Non-Autoregressive System-1 Structured Decision Model (`9.4B` Parameters)**

[![Architecture](https://img.shields.io/badge/Backbone-Qwen3.5--9B_Hybrid_(9.4B_Total)-3b82f6?style=flat-square)](#architecture)
[![Decision Head](https://img.shields.io/badge/Head-JointSchemaHead_(116M)-10b981?style=flat-square)](#architecture)
[![Decision Index](https://img.shields.io/badge/Decision_Index_v0.2.1-80.55%25_(%232_Overall_%C2%B7_%231_9B)-f59e0b?style=flat-square)](#benchmark-evaluation)
[![License](https://img.shields.io/badge/License-Apache_2.0-6366f1?style=flat-square)](./LICENSE)

</div>

**FastBrain-9B** (`9.4B` total parameters: `Qwen3.5-9B` hybrid backbone + `116M` `JointSchemaHead`) is a single-forward-pass, non-autoregressive **System-1 structured decision model**.

Given an arbitrary JSON state and one or more structured decision schemas (`choice`, `noul` binary verification, or `score` ranking), **FastBrain-9B** evaluates all candidate options and questions **jointly in a single forward pass (`0` generated tokens)**—delivering calibrated option probabilities, tool-call selections, and open-set abstention decisions at encoder-level latency.

All fine-tuned backbone updates (`104` projection layers across all `32` hybrid blocks) are merged directly into the `9B` backbone weights (`model-00001-of-00004.safetensors` .. `model-00004-of-00004.safetensors`), and all `122` decision-head tensors are packaged in `joint_head.safetensors`. **No external LoRA or PEFT adapter is required at inference time.**

---

## Benchmark Evaluation & Parameter Scale Comparison

Evaluated on the official **Cloudflare Decision Index (v0.2.1)** across **10 benchmarks (22,427 evaluation requests)**, **FastBrain-9B (`9.4B`)** achieves an overall Macro Mean of **`80.55%`**:
- **Ranks `#1` across all `<=9B` models on `10 / 10` benchmarks**, outperforming the same-size base model `Clef-flash` (`9.4B`, `74.96%`) by **`+5.59 pp`**, `Kev 9B` (`9B`, `76.51%`) by **`+4.04 pp`**, and `Laya` (`0.4B`, `58.09%`) by **`+22.46 pp`**.
- **Outperforms `3x`-larger `27B`-class models**: surpasses `Jev` (`27B`, `80.32%`) and `DiffusionGemma Jev` (`26B-A4B`, `69.14%`) in 10-benchmark Macro Mean, and outperforms the `3x`-larger flagship `Clef` (`27.4B`, `81.81%`) on **5 out of 10 benchmarks** (`BFCL`, `API-Bank`, `CLINC150+OOS`, `Home appliances`, and `When2Call`).

![FastBrain-9B Overview](./assets/benchmark_overview.png)

### Full 10-Benchmark Evaluation Matrix (Annotated by Model Parameter Scale)

![FastBrain-9B Scoreboard](./assets/benchmark_scoreboard.png)

### Fine-Grained Decision Calibration & Open-Set Reliability

![FastBrain-9B Calibration Breakdown](./assets/decision_calibration_breakdown.png)

### Summary Comparison by Model Parameter Scale

| Benchmark | Metric | `Laya`<br>*(0.4B / 421M)* | `DiffusionGemma Jev`<br>*(26B-A4B · 4B Act)* | `Clef-flash`<br>*(9.4B Base)* | `Kev 9B`<br>*(9B Dense)* | `Jev`<br>*(27B Dense)* | `Clef`<br>*(27.4B Flagship)* | **FastBrain-9B**<br>*(9.4B · Ours)* |
| :--- | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| **Parameter Scale** | **Total (Active)** | `0.42B` | `25.8B (4B MoE)` | `9.41B` | `9.0B` | `27.0B` | `27.36B` | **`9.41B`** |
| **BFCL** | Case Exact | 95.87% | 97.64% | 98.76% | 97.58% | 98.58% | 98.47% | **98.76% (#1)** |
| **ToolRet** | nDCG@10 | 51.31% | 45.28% | 67.26% | 65.72% | 65.50% | 68.18% | **67.26% (#2 · #1 9B)** |
| **API-Bank** | Accuracy | 86.22% | 85.24% | 93.11% | 89.76% | 90.35% | 91.93% | **93.70% (#1)** |
| **BANKING77** | Macro-F1 | 62.66% | 93.80% | 93.69% | 91.77% | 93.58% | 94.07% | **93.77% (#1 9B)** |
| **CLINC150+OOS** | Macro-F1 | 73.79% | 94.64% | 66.77% | 96.65% | 96.44% | 97.43% | **98.44% (#1)** |
| **Home appliances** | Case Exact | 23.86% | 68.18% | 96.59% | 88.64% | 88.64% | 96.59% | **97.73% (#1)** |
| **BRIGHT** | nDCG@10 | 25.00% | 18.69% | 38.49% | 38.53% | 43.37% | 48.03% | **38.58% (#1 9B)** |
| **Amazon ESCI** | Macro-F1 | 52.84% | 56.83% | 57.39% | 53.75% | 56.60% | 59.30% | **57.40% (#2 · #1 9B)** |
| **PhishNChips** | Accuracy | 60.45% | 55.65% | 71.95% | 72.10% | 88.20% | 91.70% | **78.75% (#1 9B)** |
| **When2Call** | Macro-F1 | 48.92% | 75.44% | 65.58% | 70.64% | 81.91% | 72.37% | **81.13% (#2 · #1 9B)** |
| **Macro Mean (10)** | **Average** | **58.09%** | **69.14%** | **74.96%** | **76.51%** | **80.32%** | **81.81%** | **80.55% (#2 · #1 9B)** |

---

## Architecture

1. **Hybrid Gated DeltaNet + Full-Attention Backbone (`9.3B` Backbone, `bfloat16`)**:
   - Processes long-context tool registries, multi-turn agent states, and candidate option spans in a single prefill pass up to `262,144` tokens (`24` linear-attention blocks + `8` full-attention blocks).
   - Jointly adapted across all `104` linear-attention (`in_proj_qkv`, `in_proj_z`, `out_proj`) and self-attention (`q_proj`, `k_proj`, `v_proj`, `o_proj`) projections and merged in-place into the backbone weights.
2. **Joint Schema Decision Head (`joint_head.safetensors`, `116M` Parameters)**:
   - Extracts question-level and option-level span representations directly from the final hidden states and scores all questions (`choice`, `noul`, `score`) simultaneously without autoregressive decoding (`output_tokens = 0`).
3. **Compile-Time AST Schema-Aware Margin Calibration (`joint_schema_model.py`)**:
   - Eliminates System-1 tool over-triggering and open-set false acceptance via compile-time schema topology analysis and post-hoc logit margin calibration.

---

## Quickstart

### 1. PyTorch / Transformers (`bfloat16` or `4-bit` / `8-bit` Quantized on Colab T4)

`TypeScript/FastBrain-9B` is a drop-in replacement for `Cloudflare/clef-flash` using the standard `joint_schema_model` runtime. It supports both full-precision `bfloat16` (`18.82 GB`, for `>=24 GB` GPUs such as L4 / RTX 4090 / PRO 6000) and mixed-precision **`load_in_4bit=True` (`~6.8 GB` VRAM, runs on a free Google Colab T4 16GB GPU)** or **`load_in_8bit=True` (`~11.2 GB` VRAM)** while keeping the `116M` `JointSchemaHead` in native `bfloat16`/`float16`:

```python
from huggingface_hub import snapshot_download
import sys
import torch

model_dir = snapshot_download("TypeScript/FastBrain-9B", ignore_patterns=["*.gguf"])
sys.path.insert(0, model_dir)
import joint_schema_model as clef

# Full precision (24GB+ GPU: L4, A100, RTX 3090/4090, RTX PRO 6000):
model, processor = clef.load_release_model(model_dir, device="cuda", dtype=torch.bfloat16)

# Or 4-bit NF4 mixed-precision (8GB–16GB GPU, e.g., free Google Colab T4):
# model, processor = clef.load_release_model(model_dir, device="cuda", dtype=torch.float16, load_in_4bit=True)

response = clef.systemone(model, processor, {
    "model": "TypeScript/FastBrain-9B",
    "state": {
        "query": "Can you check my account balance?",
        "tools": ["get_balance", "transfer_funds"],
    },
    "questions": {
        "q": {
            "type": "choice",
            "query": "Select the best next action.",
            "criteria": {
                "a": "Call <tool_call>{\"name\": \"get_balance\", \"arguments\": {}}</tool_call>",
                "b": "Could you please provide your account ID or authentication token first?",
                "c": "I am unable to access personal banking records.",
            },
        },
    },
})
print(response)
```

### 2. `llama.cpp` GGUF (`Q8_0` / `Q4_K_M` / `BF16`) via `/v1/systemone`

This repository provides official `llama.cpp` (**`b11430`+**, `ClefModel` architecture) GGUF files that bundle both the `Qwen3.5-9B` hybrid backbone and the `116M` `JointSchemaHead` (`dec.*` decision-head tensors, `ssm_` linear-attention states, and `attn_` projections preserved at `Q8_0` precision even in the `Q4_K_M` build).

> [!IMPORTANT]
> **Required `llama.cpp` Version & `ClefModel` Batch Flags:**
> 1. **Minimum version**: Use **`llama.cpp` release `b11430` or newer** (which includes native `ClefModel` / `LLM_ARCH_CLEF` graph construction and the `/v1/systemone` HTTP endpoint). Standard Ollama or older `llama.cpp` builds without `ClefModel` support cannot execute `dec.*` decision-head tensors.
> 2. **Set `-b` and `-ub` to match your maximum prompt length (e.g., `-b 16384 -ub 16384 --no-cache-prompt` or `-b 4096 -ub 4096 --no-cache-prompt`)**:
>    In `llama-server`, `ClefModel` computes cross-attention over `<|question_*|>` and `<|option_*|>` spans within a single physical micro-batch (`ubatch`, default `512`). Setting `-ub` equal to `-b` ensures long prompts (>512 tokens) are evaluated in a single physical `ubatch` without splitting option spans across micro-batches, and `--no-cache-prompt` prevents prefix-cache skipping over decision markers.

#### Step 1: Download Prebuilt `llama-server` (`b11430`+) and GGUF Weights

```bash
# 1. Download prebuilt llama.cpp (b11430+) with ClefModel & /v1/systemone support
#    - Linux CUDA 13.x (Colab / Ubuntu): llama-b11430-bin-ubuntu-cuda-13.4-x64.tar.gz
#    - Linux CUDA 12.x:                  llama-b11430-bin-ubuntu-cuda-12.8-x64.tar.gz
#    - macOS Apple Silicon (M1/M2/M3/M4): llama-b11430-bin-macos-arm64.tar.gz
curl -sL https://github.com/ggml-org/llama.cpp/releases/download/b11430/llama-b11430-bin-ubuntu-cuda-13.4-x64.tar.gz | tar -xz

# 2. Download FastBrain-9B GGUF (choose Q8_0 [9.66 GB, lossless] or Q4_K_M [6.73 GB, 8GB–16GB GPUs])
pip install -U "huggingface_hub[hf_transfer]"
HF_HUB_ENABLE_HF_TRANSFER=1 hf download TypeScript/FastBrain-9B FastBrain-9B-Q8_0.gguf --local-dir .
```

#### Step 2: Launch `llama-server` on the `/v1/systemone` Endpoint

```bash
# On Linux / Google Colab, ensure CUDA driver libraries are in LD_LIBRARY_PATH:
export LD_LIBRARY_PATH="$(pwd)/llama-b11430:/usr/lib64-nvidia:/usr/local/cuda/lib64:${LD_LIBRARY_PATH}"

# Launch llama-server (24GB+ GPU / 16K context):
./llama-b11430/llama-server \
  -m ./FastBrain-9B-Q8_0.gguf \
  -ngl 99 -c 16384 -b 16384 -ub 16384 -np 1 -fa on --no-cache-prompt \
  --port 8080

# Or on a 12GB–16GB GPU (e.g., Colab Free Tesla T4, 4K context with cuda-12.8):
# pip install nvidia-cuda-runtime-cu12 nvidia-cublas-cu12
# export LD_LIBRARY_PATH="$(pwd)/llama-b11430:$(python3 -c 'import glob;print(":".join(glob.glob("/usr/local/lib/python*/dist-packages/nvidia/*/lib")))'):/usr/lib64-nvidia:${LD_LIBRARY_PATH}"
# ./llama-b11430/llama-server \
#   -m ./FastBrain-9B-Q4_K_M.gguf \
#   -ngl 99 -c 4096 -b 4096 -ub 4096 -np 1 -fa on --no-cache-prompt \
#   --port 8080

# Optional: For multimodal image/screenshot states, attach the vision projector:
#   --mmproj ./mmproj-FastBrain-9B-Q8_0.gguf
```

#### Step 3: Query `/v1/systemone` (`curl` or Python — Single Forward Pass, `0` Output Tokens)

```bash
curl http://127.0.0.1:8080/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{
    "model": "FastBrain-9B",
    "state": {
      "ticket_id": "ORD-99281",
      "customer_message": "I was charged twice for my subscription last week and nobody has replied to my ticket."
    },
    "questions": {
      "route": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
          "billing": "Payments, refunds, duplicate charges, or invoices",
          "technical": "Bugs, crashes, or service outages",
          "account": "Password reset or login issues"
        }
      },
      "urgent": {
        "type": "noul",
        "instructions": "Is the customer reporting a financial or urgent issue?"
      }
    }
  }'
```

**Python Client Example (`requests` / `urllib`)**:

```python
import requests

resp = requests.post(
    "http://127.0.0.1:8080/v1/systemone",
    json={
        "model": "FastBrain-9B",
        "state": {
            "query": "Can you check the live wire transfer status for my corporate checking account right now?",
            "available_tools": ["get_account_balance", "list_recent_cards"],
        },
        "questions": {
            "action": {
                "type": "choice",
                "instructions": "Select the best assistant response.",
                "criteria": {
                    "direct_answer": "Your wire transfer has been completed.",
                    "tool_call": "<tool_call>{\"name\": \"get_account_balance\", \"arguments\": {}}</tool_call>",
                    "clarify": "Could you please provide your corporate account ID and wire reference number?",
                    "abstain": "I cannot check wire transfer status because no wire-transfer tool is available.",
                },
            }
        },
    },
    timeout=30,
).json()

print(resp["choices"]["action"])  # Calibrated probability distribution over all options
print(resp["usage"])              # {'input_tokens': ..., 'output_tokens': 0, 'total_tokens': ...}
```

---

## GGUF Quantization Accuracy Retention & Speed Benchmarks

### 1. Decision Index 10-Benchmark Test-Set Accuracy Retention (`BF16` vs. `Q8_0` vs. `Q4_K_M`)

Evaluated across a stratified 10-benchmark Decision Index test suite (**`203` multi-constraint requests / `1,473` scored schema questions** spanning `choice`, `noul`, and `score` heads), `FastBrain-9B-Q8_0.gguf` (`9.66 GB`) is **virtually lossless (`99.86%` Top-1 decision agreement with native `bfloat16`, `-0.03 pp` score delta)**, while `FastBrain-9B-Q4_K_M.gguf` (`6.73 GB`, with `dec.*` head preserved in `Q8_0`) retains **`98.78%` Top-1 agreement** and **`-0.20 pp`** calibrated 10-task mean:

| Benchmark / Metric | Native PyTorch `bfloat16` (`18.82 GB`) | `FastBrain-9B-Q8_0.gguf` (`9.66 GB`) | `FastBrain-9B-Q4_K_M.gguf` (`6.73 GB`, `Q8_0` Head) |
| :--- | :---: | :---: | :---: |
| **Top-1 Decision Agreement vs. `bfloat16`** (`1,473` questions) | `100.00%` *(Reference)* | **`99.86%` (`1,471 / 1,473` identical)** | **`98.78%` (`1,455 / 1,473` identical)** |
| **Probability Mean Absolute Error (`|p_quant - p_bf16|`)** | `0.00000` | **`0.00203` (`0.20%`)** | **`0.00387` (`0.39%`)** |
| **10-Benchmark Mean (Single-Checkpoint Raw)** | **`82.37%`** | **`82.34%` (`-0.03 pp`)** | **`82.56%` (`+0.19 pp`)** |
| **10-Benchmark Mean (With Margin Calibration)** | **`83.68%`** | **`83.65%` (`-0.03 pp`)** | **`83.48%` (`-0.20 pp`)** |
| ├─ **1. BFCL** *(AST Tool Selection)* | `100.00%` | **`100.00%`** (`0.00`) | **`100.00%`** (`0.00`) |
| ├─ **2. ToolRet** *(20-Candidate Tool Reranking, `score`)* | `73.15%` | **`73.70%`** (`+0.55`) | **`75.60%`** (`+2.45`) |
| ├─ **3. API-Bank** *(Multi-Step Stateful Tool Calling)* | `95.00%` | **`95.00%`** (`0.00`) | **`95.00%`** (`0.00`) |
| ├─ **4. BANKING77** *(77-Way Fine-Grained Intent)* | `100.00%` | **`100.00%`** (`0.00`) | **`100.00%`** (`0.00`) |
| ├─ **5. CLINC150+OOS** *(Intent + Open-Set `noul` Abstention)* | `93.94%` | **`93.94%`** (`0.00`) | **`93.94%`** (`0.00`) |
| ├─ **6. Home Appliances** *(Embedded Device Control)* | `90.00%` | **`90.00%`** (`0.00`) | **`90.00%`** (`0.00`) |
| ├─ **7. BRIGHT** *(Reasoning-Intensive Document Ranking)* | `41.93%` | **`41.00%`** (`-0.93`) | **`40.31%`** (`-1.62`) |
| ├─ **8. Amazon ESCI** *(Fine-Grained E-Commerce Relevance)* | `73.89%` | **`73.89%`** (`0.00`) | **`75.79%`** (`+1.90`) |
| ├─ **9. PhishNChips** *(9-Probe Joint Security Inspection)* | `85.00%` | **`85.00%`** (`0.00`) | **`85.00%`** (`0.00`) |
| └─ **10. When2Call** *(4-Way Tool Governance)* | `75.00%` *(Raw `70.83%`)* | **`75.00%`** *(Raw `70.83%`, `0.00`)* | **`72.92%`** *(Raw `70.83%`, `-2.08`)* |

### 2. Measured Hardware Footprint & Inference Speed (`llama-server b11430` CUDA 13.4)

Because `FastBrain-9B` scores all schema questions and candidate options in a **single non-autoregressive prefill forward pass (`output_tokens = 0`)**, `llama-server` achieves **8,900–12,100+ input tokens/sec** (`31–109 ms` end-to-end HTTP latency) on server GPUs:

| Workload / Metric | `Q8_0.gguf`<br>*(RTX PRO 6000)* | `Q4_K_M.gguf`<br>*(RTX PRO 6000)* | `Q8_0.gguf`<br>*(Colab Free T4 16GB)* | `Q4_K_M.gguf`<br>*(Colab Free T4 16GB)* | Native `bfloat16`<br>*(RTX PRO 6000)* |
| :--- | :---: | :---: | :---: | :---: | :---: |
| **Model Load Time** | **`2.56 s`** | **`2.00 s`** | **`5.43 s`** | **`5.10 s`** | `~11.5 s` |
| **Peak GPU VRAM** | **`12.55 GiB`** *(16K ctx)* | **`10.30 GiB`** *(16K ctx)* | **`9.03 GiB` (`9,245 MiB`)** *(4K ctx)* | **`6.78 GiB` (`6,943 MiB`)** *(4K ctx)* | `20.3 – 24.7 GiB` |
| **Workload 1: Intent Routing + OOS** (`293 tok`, `K=2`) | **`32.82 ms`** (`8,928 tok/s`) | **`31.69 ms`** (`9,245 tok/s`) | **`349.32 ms`** (`839 tok/s`) | **`342.98 ms`** (`854 tok/s`) | `~35.0 ms` |
| **Workload 2: Tool Governance (`When2Call`)** (`1,068 tok`, `K=1`) | **`87.78 ms`** (`12,167 tok/s`) | **`89.34 ms`** (`11,955 tok/s`) | **`1,149.81 ms`** (`929 tok/s`) | **`1,154.31 ms`** (`925 tok/s`) | `~92.0 ms` |
| **Workload 3: Security Multi-Probe (`PhishNChips`)** (`1,303 tok`, `K=9`) | **`108.55 ms`** (`12,004 tok/s`) | **`109.97 ms`** (`11,849 tok/s`) | **`1,430.81 ms`** (`911 tok/s`) | **`1,440.52 ms`** (`905 tok/s`) | `~115.0 ms` |

| Format / Precision Tier | Weight Footprint | Peak VRAM (`4K`–`16K` ctx) | Minimum Recommended Hardware |
| :--- | :---: | :---: | :--- |
| **`FastBrain-9B-Q4_K_M.gguf`** *(4-bit + `Q8_0` Head)* | `6.73 GB` | `~7.8 – 10.3 GiB` | **8GB–12GB GPU / Colab Free T4 (16GB) / Mac M1–M4** |
| **PyTorch `load_in_4bit=True`** *(NF4 + `fp16` Head)* | `~5.5 GB` | `~6.2 – 6.8 GiB` | **Colab Free T4 (16GB) / RTX 3060–4060 (8GB+)** |
| **`FastBrain-9B-Q8_0.gguf`** *(Full 8-bit GGUF)* | `9.66 GB` | `~10.8 – 12.6 GiB` | **12GB–16GB GPU / Colab T4 / Mac 16GB+** |
| **PyTorch `load_in_8bit=True`** *(INT8 + `bf16` Head)* | `~9.6 GB` | `~11.2 GiB` | **Colab T4 (16GB) / RTX 4070 Ti (12GB+)** |
| **Native `bfloat16` (`safetensors` / `BF16.gguf`)** | `18.82 GB` (`18.16 GB`) | `20.3 – 24.7 GiB` | **Colab L4 (24GB) / RTX 3090–4090 / PRO 6000 (96GB)** |

---

## Repository Structure

| File | Size | Description |
| :--- | :---: | :--- |
| `model-00001-of-00004.safetensors` .. `00004` | `18.59 GB` | Full `9B` backbone weights (`bfloat16`, in-place merged, no LoRA required) |
| `joint_head.safetensors` & `joint_head_config.json` | `232 MB` | `116M`-parameter `JointSchemaHead` weights (`122` tensors) and configuration |
| `FastBrain-9B-Q4_K_M.gguf` | `6.73 GB` | Recommended 4-bit GGUF (`dec.*` decision head, `ssm_`, and `attn_` preserved at `Q8_0`) |
| `FastBrain-9B-Q8_0.gguf` | `9.66 GB` | High-precision 8-bit GGUF (`ClefModel` architecture for `llama-server`) |
| `FastBrain-9B-BF16.gguf` | `18.16 GB` | Full-precision `bfloat16` GGUF (`ClefModel` architecture for `llama-server`) |
| `mmproj-FastBrain-9B-Q8_0.gguf` & `BF16.gguf` | `624 MB` / `922 MB` | Multimodal vision projector GGUF files for image/screenshot states |
| `joint_schema_model.py` | `25 KB` | `ClefModel`, `JointSchemaHead`, `load_release_model(load_in_4bit/8bit)`, and `systemone()` runtime |
| `config.json`, `tokenizer.json`, `processor_config.json` | `11 MB` | Standard Hugging Face model, tokenizer, and processor configurations |
| `assets/*.png` | `1.2 MB` | High-resolution technical benchmark and calibration charts |