Language 涓枃|English
Model Details
This model is an NVFP4A16 quantized version of Qwen3.5-35B-A3B generated with llm-compressor. Please follow the license of the original model.
This model uses Instruct mode by default. To enable thinking mode, please change {%- set enable_thinking = false %} to true in chat_template.jinja, and make sure to add --reasoning-parser qwen3 to the model launch command.Like thinking-mode.
Quantization Strategy
| Layer Type | Bits | Notes |
|---|---|---|
lm_head |
16-bit | Kept in original precision to preserve final token prediction quality and avoid extra degradation at output projection |
visual.* |
16-bit | Kept in original precision to better preserve visual feature extraction quality and reduce multimodal degradation |
model.visual.* |
16-bit | Kept in original precision to maintain stable vision backbone and vision-language feature processing |
mlp.gate |
16-bit | Kept in original precision to preserve gating accuracy and reduce quality loss from gate quantization |
embed_tokens |
16-bit | Kept in original precision to maintain input embedding fidelity and reduce quantization error accumulation from the first layer |
shared_expert_gate |
16-bit | Kept in original precision to preserve expert routing stability and reduce quality loss from gate quantization |
linear_attn.* |
16-bit | Kept in original precision to maintain attention accuracy and stabilize long-context / sequence modeling behavior |
mtp.fc.* |
16-bit | Kept in original precision to preserve speculative / MTP prediction quality and decoding stability |
Model Size
| Model | Bits | Model Size | Coding: LiveCodeBench v6 | Multilingual: MMLU-ProX | OpenClaw: PinchBench |
|---|---|---|---|---|---|
| Qwen3.5-35B-A3B | Original BF16 | ~72GB | 0.75 | 0.81 | 0.81 |
| NVFP4A16 | ~25GB (65% reduction 鈫撯啌) | 0.73 | 0.71 | 0.79 | |
| mixed INT4 | ~22GB | 0.58 | 0.69 | 0.71 | |
| Qwen3-Coder-30B-A3B | W4A16 | ~17GB | 0.51 | 0.57 | 0.37 |
Quickstart
vLLM Usage
vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.
Directly talk to the model
import argparse
import atexit
import json
import os
import shutil
import subprocess
import sys
import time
import urllib.error
import urllib.request
# ---------------------------
# User-facing configuration
# ---------------------------
DEFAULTS = {
"model": "YCWTG/Qwen3.5-35B-A3B-Instruct-NVFP4A16",
"served_model_name": "YCWTG/Qwen3.5-35B-A3B-Instruct-NVFP4A16",
"host": "localhost",
"port": 8000,
"max_model_len": 143616,
"enable_auto_tool_choice": True,
"tool_call_parser": "qwen3_coder",
"attention_backend": "FLASHINFER",
"max_num_seqs": 1,
"allowed_local_media_path": None,
"trust_remote_code": True,
"language_model_only": False,
}
RUNTIME = {
"gpu_memory_utilization": 0.97,
"startup_timeout_sec": 180,
"healthcheck_timeout_sec": 3,
"healthcheck_interval_sec": 1,
"chat_timeout_sec": 600,
}
SERVE_VALUE_OPTIONS = (
("--served-model-name", "served_model_name"),
("--host", "host"),
("--port", "port"),
("--max-model-len", "max_model_len"),
("--tool-call-parser", "tool_call_parser"),
("--attention-backend", "attention_backend"),
("--max-num-seqs", "max_num_seqs"),
)
CLIENT_VALUE_OPTIONS = (
("--model", "model"),
*SERVE_VALUE_OPTIONS,
)
SERVE_BOOL_OPTIONS = (
("--enable-auto-tool-choice", "enable_auto_tool_choice"),
("--trust-remote-code", "trust_remote_code"),
("--language-model-only", "language_model_only"),
)
CLIENT_BOOL_OPTIONS = (
("--language-model-only", "--no-language-model-only", "language_model_only"),
("--enable-auto-tool-choice", "--no-enable-auto-tool-choice", "enable_auto_tool_choice"),
("--trust-remote-code", "--no-trust-remote-code", "trust_remote_code"),
)
def append_value_options(cmd, args, options):
for flag, attr in options:
cmd.extend([flag, str(getattr(args, attr))])
def append_true_bool_options(cmd, args, options):
for flag, attr in options:
if getattr(args, attr):
cmd.append(flag)
def append_boolean_optional_options(cmd, args, options):
for positive_flag, negative_flag, attr in options:
cmd.append(positive_flag if getattr(args, attr) else negative_flag)
def append_optional_value_option(cmd, args, flag, attr):
value = getattr(args, attr)
if value is None:
return
if isinstance(value, str) and not value.strip():
return
cmd.extend([flag, str(value)])
def multiline_input():
print('User (type "END" on a single line to send, "exit" to quit):')
lines = []
while True:
line = input()
text = line.strip()
if text.lower() in {"exit", "quit"}:
return None
if text == "END":
break
lines.append(line)
return "\n".join(lines)
def resolve_client_host(host):
return "127.0.0.1" if host in {"0.0.0.0", "::"} else host
def launch_vllm(args):
cmd = ["vllm", "serve", args.model]
append_value_options(cmd, args, SERVE_VALUE_OPTIONS)
append_optional_value_option(cmd, args, "--allowed-local-media-path", "allowed_local_media_path")
cmd.extend(
[
"--gpu-memory-utilization",
str(RUNTIME["gpu_memory_utilization"]),
]
)
append_true_bool_options(cmd, args, SERVE_BOOL_OPTIONS)
print("Launching vLLM:")
print(" ".join(cmd))
try:
return subprocess.Popen(cmd)
except FileNotFoundError as e:
raise RuntimeError("vllm command not found. Activate an environment that has vllm installed.") from e
def stop_vllm(proc):
if proc and proc.poll() is None:
proc.terminate()
try:
proc.wait(timeout=10)
except subprocess.TimeoutExpired:
proc.kill()
def wait_vllm_ready(base_url, timeout_sec=RUNTIME["startup_timeout_sec"]):
deadline = time.time() + timeout_sec
url = f"{base_url}/v1/models"
req = urllib.request.Request(url=url)
while time.time() < deadline:
try:
with urllib.request.urlopen(req, timeout=RUNTIME["healthcheck_timeout_sec"]) as resp:
if resp.status == 200:
return True
except urllib.error.URLError:
pass
time.sleep(RUNTIME["healthcheck_interval_sec"])
return False
def chat_once(base_url, model_name, messages):
payload = {"model": model_name, "messages": messages}
req = urllib.request.Request(
url=f"{base_url}/v1/chat/completions",
data=json.dumps(payload, ensure_ascii=False).encode("utf-8"),
headers={"Content-Type": "application/json"},
method="POST",
)
with urllib.request.urlopen(req, timeout=RUNTIME["chat_timeout_sec"]) as resp:
data = json.loads(resp.read().decode("utf-8"))
return data["choices"][0]["message"]
def chat_loop(base_url, model_name):
print("\n===== Chat Started =====\n")
messages = []
while True:
user_text = multiline_input()
if user_text is None:
break
messages.append({"role": "user", "content": user_text})
try:
assistant_msg = chat_once(base_url, model_name, messages)
except Exception as e:
print(f"\nRequest failed: {e}\n")
messages.pop()
continue
content = assistant_msg.get("content")
tool_calls = assistant_msg.get("tool_calls")
if content:
print(f"\nAssistant:\n{content}\n")
elif tool_calls:
print("\nAssistant(tool_calls):")
print(json.dumps(tool_calls, ensure_ascii=False, indent=2))
print()
else:
print("\nAssistant:\n(empty response)\n")
normalized_msg = {"role": "assistant", "content": content or ""}
if tool_calls:
normalized_msg["tool_calls"] = tool_calls
messages.append(normalized_msg)
def build_client_command(args):
cmd = [sys.executable, os.path.abspath(__file__), "--_client"]
append_value_options(cmd, args, CLIENT_VALUE_OPTIONS)
append_boolean_optional_options(cmd, args, CLIENT_BOOL_OPTIONS)
return cmd
def spawn_chat_terminal(args):
client_cmd = build_client_command(args)
terminal_cmd = None
if os.name == "nt":
# Open a new cmd window on Windows and keep it alive for interactive chat.
terminal_cmd = [
"cmd",
"/c",
"start",
"",
"cmd",
"/k",
subprocess.list2cmdline(client_cmd),
]
elif shutil.which("gnome-terminal"):
terminal_cmd = ["gnome-terminal", "--", *client_cmd]
elif shutil.which("x-terminal-emulator"):
terminal_cmd = ["x-terminal-emulator", "-e", *client_cmd]
if not terminal_cmd:
return False
try:
subprocess.Popen(terminal_cmd)
return True
except Exception as e:
print(f"Failed to open a new terminal automatically: {e}")
return False
def parse_args():
parser = argparse.ArgumentParser(description="Minimal local vLLM chat script")
parser.add_argument("--_client", action="store_true", help=argparse.SUPPRESS)
parser.add_argument("--model", default=DEFAULTS["model"])
parser.add_argument(
"--served-model-name",
default=DEFAULTS["served_model_name"],
)
parser.add_argument("--host", default=DEFAULTS["host"])
parser.add_argument("--port", type=int, default=DEFAULTS["port"])
parser.add_argument("--max-model-len", type=int, default=DEFAULTS["max_model_len"])
parser.add_argument(
"--max-num-seqs",
"--max_num_seqs",
dest="max_num_seqs",
type=int,
default=DEFAULTS["max_num_seqs"],
)
parser.add_argument(
"--enable-auto-tool-choice",
action=argparse.BooleanOptionalAction,
default=DEFAULTS["enable_auto_tool_choice"],
)
parser.add_argument(
"--allowed-local-media-path",
default=DEFAULTS["allowed_local_media_path"],
help="Optional local media path. Leave empty to disable.",
)
parser.add_argument("--tool-call-parser", default=DEFAULTS["tool_call_parser"])
parser.add_argument("--attention-backend", default=DEFAULTS["attention_backend"])
parser.add_argument(
"--trust-remote-code",
action=argparse.BooleanOptionalAction,
default=DEFAULTS["trust_remote_code"],
)
parser.add_argument(
"--language-model-only",
action=argparse.BooleanOptionalAction,
default=DEFAULTS["language_model_only"],
)
return parser.parse_args()
def main():
args = parse_args()
base_url = f"http://{resolve_client_host(args.host)}:{args.port}"
if args._client:
chat_loop(base_url, args.served_model_name)
return
proc = launch_vllm(args)
atexit.register(stop_vllm, proc)
print(f"Waiting for service to become ready: {base_url}")
if not wait_vllm_ready(base_url):
print("vLLM startup timed out. Check server logs above.")
stop_vllm(proc)
sys.exit(1)
if spawn_chat_terminal(args):
print("Model is ready. Opened a new terminal for chat; this terminal keeps server logs.")
print("Press Ctrl+C here to stop vLLM.")
try:
proc.wait()
except KeyboardInterrupt:
print("\nInterrupted. Stopping vLLM...")
else:
print("No supported terminal found. Falling back to chat in this terminal.")
chat_loop(base_url, args.served_model_name)
if __name__ == "__main__":
main()
Directly use the OpenAPI
Instruct Mode
vllm serve --model YCWTG/Qwen3.5-35B-A3B-Instruct-NVFP4A16 --served-model-name YCWTG/Qwen3.5-35B-A3B-Instruct-NVFP4A16 --host localhost --port 8000 --max-model-len 143616 --enable-auto-tool-choice --tool-call-parser qwen3_coder --trust-remote-code --gpu-memory-utilization 0.97 --attention-backend FLASHINFER --max_num_seqs 1 --allowed-local-media-path /home/ycwtg/image
Thinking Mode
vllm serve --model YCWTG/Qwen3.5-35B-A3B-Instruct-NVFP4A16 --served-model-name YCWTG/Qwen3.5-35B-A3B-Instruct-NVFP4A16 --host localhost --port 8000 --max-model-len 143616 --enable-auto-tool-choice --tool-call-parser qwen3_coder --trust-remote-code --gpu-memory-utilization 0.97 --attention-backend FLASHINFER --max_num_seqs 1 --allowed-local-media-path /home/ycwtg/image --reasoning-parser qwen3
See its documentation for more details.
The following will create API endpoints at http://localhost:8000/v1.
Generate the Model
See code here.
Ethical Considerations and Limitations
The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs.
Therefore, before deploying any applications of the model, developers should perform safety testing.
Disclaimer
The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.
- Downloads last month
- 1,038