d3-flash-GGUF

d3-flash is an 8.39B-parameter multimodal decision model from the vLLM Semantic Router team, including a 0.46B vision encoder. It is built on Qwen3.5-9B and released under Apache-2.0. It takes text or JSON input, optionally with several images and videos, and answers multiple Choice, Yes/No and Score questions about it in one call. It returns a probability for every option and generates no text. On the official Jev Decision Index 0.3.2 text board it scores 55.8, which is 7.6 points above Decision 2.0's 9B model. It ranks first among similarly sized entries and sits on the Pareto frontier, though its lead over Bespoke Nimble 9B v3 (54.7) is narrow. It scores 66.9 on the vision board, first among the three entries of similar size, and 73.3 on the Perception Test video benchmark. On one AMD Instinct MI325X, median latency is 19.2 ms for a text request, 166 ms with an image and 371 ms with a 10-second video. vLLM Semantic Router's Engine mode with batching serves up to 95 one-question text requests per second, 1.8× AutoModel. It can be used through transformers (AutoModel or a pipeline) or served with vllm-sr serve -e vllm-sr/d3-flash.

This is a Decision Model GGUF that supports the llama.cpp v1/systemone API endpoint.

Model Files

File Name	Quant Type	File Size	File Link	Description
d3-flash.BF16.gguf	BF16	17.9 GB	Link	Full BF16 weights. Highest quality, largest file size.
d3-flash.Q3_K_L.gguf	Q3_K_L	4.93 GB	Link	Lower quality but usable, good for low RAM availability.
d3-flash.Q3_K_M.gguf	Q3_K_M	4.62 GB	Link	Low quality.
d3-flash.Q4_K_M.gguf	Q4_K_M	5.63 GB	Link	Good quality, default size for most use cases, recommended.
d3-flash.Q4_K_S.gguf	Q4_K_S	5.35 GB	Link	Slightly lower quality with more space savings, recommended.
d3-flash.Q5_K_M.gguf	Q5_K_M	6.47 GB	Link	High quality, recommended.
d3-flash.Q5_K_S.gguf	Q5_K_S	6.31 GB	Link	High quality, recommended.
d3-flash.Q6_K.gguf	Q6_K	7.36 GB	Link	Very high quality, near perfect, recommended.
d3-flash.Q8_0.gguf	Q8_0	9.53 GB	Link	Extremely high quality, generally unneeded but max available quant.
d3-flash.mmproj-bf16.gguf	mmproj-bf16	922 MB	Link	Multimodal projection file in BF16 format. Used for vision/language models.

Run with llama.cpp (SystemOne API)

1. Start the Server

Launch llama-server with the GGUF model and wait for the health check endpoint to report ready.

import subprocess, time, requests

MODEL_PATH = "/data/llama.cpp/gguf-new/d3-flash.BF16.gguf"
BASE_URL = "http://127.0.0.1:8080"

server = subprocess.Popen(
    [
        "/data/llama.cpp/build/bin/llama-server",
        "-m", MODEL_PATH,
        "--host", "127.0.0.1",
        "--port", "8080",
        "-ngl", "99",
    ],
    stdout=open("/data/llama-server.log", "w"),
    stderr=subprocess.STDOUT,
)

for _ in range(120):
    try:
        if requests.get(f"{BASE_URL}/health", timeout=2).status_code == 200:
            print("server ready")
            break
    except requests.exceptions.RequestException:
        pass
    time.sleep(2)
else:
    print("server not ready, check /data/llama-server.log")

2. Query the /v1/systemone Endpoint

Send a POST request with a state context and structured questions (choice, noul, or score).

curl -s http://127.0.0.1:8080/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{
    "state": "Customer message: I was charged twice for my order last week and nobody has replied.",
    "questions": {
      "route": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
          "billing": "payments, charges, refunds, invoices",
          "shipping": "delivery, tracking, lost or late parcels",
          "technical": "bugs, errors, login problems"
        }
      },
      "angry": {
        "type": "noul",
        "instructions": "Is the customer angry?"
      },
      "urgency": {
        "type": "score",
        "instructions": "How urgent is this?",
        "criteria": ["can wait", "this week", "today", "right now"]
      }
    }
  }' | python3 -m json.tool

3. Example Response

{
    "model": "/data/llama.cpp/gguf-new/d3-flash.BF16.gguf",
    "answers": {
        "route": {
            "type": "choice",
            "choice": "billing",
            "probabilities": {
                "billing": 0.9978382942165583,
                "shipping": 0.0012888281316195192,
                "technical": 0.0008728776518221586
            },
            "confidence": 0.9846319216110954
        },
        "angry": {
            "type": "noul",
            "noul": 0.9158028892594581
        },
        "urgency": {
            "type": "score",
            "score": 2.197597492798958,
            "legend": {
                "0": "can wait",
                "1": "this week",
                "2": "today",
                "3": "right now"
            },
            "probabilities": {
                "0": 0.03539091724580157,
                "1": 0.21014318581391891,
                "2": 0.27594338383579925,
                "3": 0.47852251310448024
            },
            "confidence": 0.16752351459947223
        }
    },
    "usage": {
        "input_tokens": 350,
        "output_tokens": 0
    }
}

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

Quantized using llama.cpp by prithivMLmods.