Image-Text-to-Text
MLX
Safetensors
qwen3_5
clairic
cloudflare
systemone
structured-output
information-extraction
multimodal
custom-code
conversational
4-bit precision
Instructions to use mlx-community/clairic-4bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use mlx-community/clairic-4bit with MLX:
# Make sure mlx-vlm is installed # pip install --upgrade mlx-vlm from mlx_vlm import load, generate from mlx_vlm.prompt_utils import apply_chat_template from mlx_vlm.utils import load_config # Load the model model, processor = load("mlx-community/clairic-4bit") config = load_config("mlx-community/clairic-4bit") # Prepare input image = ["http://images.cocodataset.org/val2017/000000039769.jpg"] prompt = "Describe this image." # Apply chat template formatted_prompt = apply_chat_template( processor, config, prompt, num_images=1 ) # Generate output output = generate(model, processor, formatted_prompt, image) print(output) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Pi
How to use mlx-community/clairic-4bit with Pi:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "mlx-community/clairic-4bit"
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "mlx-lm": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "mlx-community/clairic-4bit" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Hermes Agent
How to use mlx-community/clairic-4bit with Hermes Agent:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "mlx-community/clairic-4bit"
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default mlx-community/clairic-4bit
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use mlx-community/clairic-4bit with OpenClaw:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "mlx-community/clairic-4bit"
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "mlx-community/clairic-4bit" \ --custom-provider-id mlx-lm \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
File size: 8,714 Bytes
1a48d6b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 | """Convert a Cloudflare Clairic checkpoint (PyTorch `Qwen3_5ForConditionalGeneration` + typed head) to MLX.
python convert.py <src_checkpoint> <dst_dir> [--bits 4] [--group-size 64] [--mode affine]
[--keep in_proj_a,in_proj_b] [--high <regex> --high-bits 8] [--dtype bfloat16]
What it does:
- renames the backbone tensors to the `mlx_vlm.models.qwen3_5` layout with mlx-vlm's own `sanitize` (conv1d
axis move, patch-embed transpose, +1 on the zero-centred RMSNorm weights) and drops the unused `mtp.*` draft
head (the config sets `mtp_num_hidden_layers` to 0 and the runtime never calls it);
- stores the RMSNorm weights as exact fp32 `1 + w` (the runtime evaluates these norms in fp32);
- quantizes the language model's linear layers, token embedding and vocabulary head (`--bits 0` keeps
everything in `--dtype`); the vision tower, norms, conv1d, `A_log` / `dt_bias` and any `--keep` module-path
suffixes are never quantized;
- copies the typed head (`clair_head.safetensors`, fp32, unchanged), tokenizer, configs and licence.
The folder is not a stock mlx-lm / mlx-vlm chat model: answers come from the typed head and from
grammar-constrained decoding driven by the runtime shipped next to the weights.
"""
from __future__ import annotations
import argparse
import json
import re
import shutil
import time
from pathlib import Path
import mlx.core as mx
import mlx.nn as nn
from mlx.utils import tree_flatten
from mlx_vlm.models.qwen3_5 import Model, ModelConfig
SHARD_BYTES = 5 << 30
COPY = [
"LICENSE", "chat_template.jinja", "clair_head.safetensors", "clair_head_config.json", "generation_config.json",
"preprocessor_config.json", "tokenizer.json", "tokenizer_config.json",
]
NORM_SUFFIXES = (
".input_layernorm.weight", ".post_attention_layernorm.weight", "model.norm.weight", ".q_norm.weight", ".k_norm.weight",
)
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("src")
ap.add_argument("dst")
ap.add_argument("--bits", type=int, default=4, help="0 = no quantization")
ap.add_argument("--group-size", type=int, default=64)
ap.add_argument("--mode", default="affine")
ap.add_argument("--keep", default="in_proj_a,in_proj_b", help="comma-separated module-path suffixes to leave unquantized")
ap.add_argument("--high", default="", help="regular expression of module paths quantized with --high-bits instead")
ap.add_argument("--high-bits", type=int, default=8)
ap.add_argument("--dtype", default="bfloat16", choices=["bfloat16", "float32"])
args = ap.parse_args()
src, dst = Path(args.src), Path(args.dst)
dst.mkdir(parents=True, exist_ok=True)
dtype = getattr(mx, args.dtype)
keep = tuple(s for s in args.keep.split(",") if s)
config = json.loads((src / "config.json").read_text())
started = time.time()
# ---- load (lazily), rename and sanitize ----------------------------------------------------------
raw: dict[str, mx.array] = {}
for shard in sorted(src.glob("model*.safetensors")):
raw.update(mx.load(str(shard)))
dropped = [key for key in raw if key.startswith("mtp.")]
if "lm_head.weight" not in raw or not any(key.startswith("model.visual.") for key in raw):
raise ValueError("Expected a full Qwen3_5ForConditionalGeneration checkpoint (lm_head and vision tower)")
# Copies: mlx-vlm's sanitize shifts the norm weights in place (`value += 1.0`).
originals = {key: mx.array(value) for key, value in raw.items()
if key.startswith("model.language_model.") and (key.endswith(NORM_SUFFIXES) or "conv1d.weight" in key)}
model = Model(ModelConfig.from_dict(config))
weights = model.sanitize(raw)
weights = model.vision_tower.sanitize(weights)
# Zero-centred RMSNorm weights: the reference computes (1 + w) in fp32. mlx-vlm's sanitize adds the 1 in
# bf16, which rounds weights above 2 to 1/64. Store the exact fp32 sum instead (a few hundred small vectors).
norm_keys = set()
for key, value in originals.items():
new_key = "language_model.model." + key.removeprefix("model.language_model.")
new = weights[new_key]
if "conv1d.weight" in key:
assert new.shape == (value.shape[0], value.shape[2], value.shape[1]), (key, new.shape)
continue
exact = value.astype(mx.float32) + 1.0
# sanitize must have applied the shift itself (within bf16 rounding); otherwise its detection changed.
assert mx.abs(new.astype(mx.float32) - exact).max().item() <= 0.04, f"norm weight {key} was not shifted by +1"
weights[new_key] = exact
norm_keys.add(new_key)
print(f"sanitized {len(weights)} tensors ({len(dropped)} mtp tensors dropped); "
f"{len(norm_keys)} RMSNorm weights stored as exact fp32 (1 + w)")
def cast(key: str, value: mx.array) -> mx.array:
if not mx.issubdtype(value.dtype, mx.floating) or key in norm_keys:
return value
return value.astype(dtype)
weights = {key: cast(key, value) for key, value in weights.items()}
model.load_weights(list(weights.items()), strict=True)
# ---- quantize ------------------------------------------------------------------------------------
quantized: list[str] = []
skipped: list[str] = []
raised: dict[str, dict] = {}
high = re.compile(args.high) if args.high else None
if args.bits:
def predicate(path: str, module: nn.Module) -> bool | dict:
if not hasattr(module, "to_quantized"):
return False
if path.startswith("vision_tower") or path.endswith(keep or ("\0",)):
skipped.append(path)
return False
if module.weight.shape[-1] % args.group_size:
skipped.append(path)
return False
quantized.append(path)
settings = {"bits": args.bits, "group_size": args.group_size, "mode": args.mode}
if high is not None and high.search(path):
settings["bits"] = args.high_bits
raised[path] = settings
return settings
nn.quantize(model, class_predicate=predicate)
# The standard MLX layout: global settings, then per-module entries: `false` for language-model modules
# left unquantized, and the settings of modules quantized with a different bit width.
config["quantization"] = {"group_size": args.group_size, "bits": args.bits, "mode": args.mode,
**{path: False for path in sorted(skipped) if path.startswith("language_model")},
**{path: raised[path] for path in sorted(raised)}}
config["quantization_config"] = config["quantization"]
else:
config.pop("quantization", None)
config.pop("quantization_config", None)
# ---- save shards (evaluating one shard at a time) -------------------------------------------------
for old in dst.glob("model*.safetensors*"):
old.unlink()
params = dict(tree_flatten(model.parameters()))
shards: list[dict[str, mx.array]] = [{}]
size = 0
for key in sorted(params):
nbytes = params[key].nbytes
if size and size + nbytes > SHARD_BYTES:
shards.append({})
size = 0
shards[-1][key] = params[key]
size += nbytes
index = {"metadata": {"format": "mlx", "total_size": sum(v.nbytes for v in params.values()),
"total_parameters": sum(v.size for v in params.values())}, "weight_map": {}}
for number, shard in enumerate(shards, 1):
name = f"model-{number:05d}-of-{len(shards):05d}.safetensors"
mx.eval(*shard.values())
mx.save_safetensors(str(dst / name), shard, metadata={"format": "mlx"})
for key in shard:
index["weight_map"][key] = name
print(f" {name}: {len(shard)} tensors, {sum(v.nbytes for v in shard.values()) / 1e9:.2f} GB", flush=True)
for key in shard: # release
params[key] = None
shard.clear()
mx.clear_cache()
(dst / "model.safetensors.index.json").write_text(json.dumps(index, indent=2))
(dst / "config.json").write_text(json.dumps(config, indent=2))
for name in COPY:
shutil.copy2(src / name, dst / name)
total = sum(f.stat().st_size for f in dst.glob("model-*.safetensors"))
print(f"wrote {dst} in {time.time() - started:.0f}s: {total / 1e9:.2f} GB of backbone weights, "
f"{len(quantized)} modules quantized ({len(raised)} of them at {args.high_bits} bits), "
f"{len(skipped)} quantizable modules kept in {args.dtype}")
if __name__ == "__main__":
main()
|