Text Generation
MLX
Safetensors
qwen3_5
apple-silicon
macos
m1
m2
fp16
speculative-decoding
multi-token-prediction
qwen
qwen3.8
mtp
mtplx
local-ai
coding
conversational
4-bit precision
Instructions to use Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16") prompt = "Write a story about Einstein" messages = [{"role": "user", "content": prompt}] prompt = tokenizer.apply_chat_template( messages, add_generation_prompt=True ) text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Pi
How to use Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 with Pi:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16"
Configure the model in Pi
# Install Pi: npm install -g @mariozechner/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "mlx-lm": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16" } ] } } }Run Pi
# Start Pi in your project directory: pi
- MLX LM
How to use Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Interactive chat REPL mlx_lm.chat --model "Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16"
Run an OpenAI-compatible server
# Install MLX LM uv tool install mlx-lm # Start the server mlx_lm.server --model "Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16" # Calling the OpenAI-compatible server with curl curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16", "messages": [ {"role": "user", "content": "Hello"} ] }' - Hermes Agent
How to use Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 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 "Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16"
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 Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16 with OpenClaw:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16"
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 "Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16" \ --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"
Quantized MTP draft head (INT4/g64 affine, all 8 head matrices)
Browse filesmtp.safetensors: 239 MB (was 849 MB). Trunk weights are
unchanged. Acceptance re-verified flat-or-better against the
previous head before publishing; mtplx_runtime.json carries the
fresh fingerprint-bound verification rows and config.json declares
the prequantized head layout (loadable by MTPLX >= 2.0.1;
this model family needs >= 2.7.0).
- config.json +9 -1
- mtp.safetensors +2 -2
- mtplx_runtime.json +47 -35
config.json
CHANGED
|
@@ -912,5 +912,13 @@
|
|
| 912 |
"temporal_patch_size": 2
|
| 913 |
},
|
| 914 |
"vision_end_token_id": 248054,
|
| 915 |
-
"vision_start_token_id": 248053
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 916 |
}
|
|
|
|
| 912 |
"temporal_patch_size": 2
|
| 913 |
},
|
| 914 |
"vision_end_token_id": 248054,
|
| 915 |
+
"vision_start_token_id": 248053,
|
| 916 |
+
"mtplx_mtp_quantization": {
|
| 917 |
+
"bits": 4,
|
| 918 |
+
"group_size": 64,
|
| 919 |
+
"mode": "affine",
|
| 920 |
+
"policy": "all",
|
| 921 |
+
"prequantized": true,
|
| 922 |
+
"description": "All 8 MTP draft-head matrices (fc + attention q/k/v/o + MLP gate/up/down) packed MLX INT4/g64 affine from the released sidecar; head norms keep the pack's float dtype. Verified flat-or-better acceptance vs the unquantized head before publishing."
|
| 923 |
+
}
|
| 924 |
}
|
mtp.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f803737b0e64bcce54fcfa22f2c7ee342ad37192b8e2a3299fc940387a828c9d
|
| 3 |
+
size 238934226
|
mtplx_runtime.json
CHANGED
|
@@ -91,7 +91,18 @@
|
|
| 91 |
"recommended_profile_note": "Restamped sustained->turbo 2026-08-14 23:50 (mirrors the OQ restamp): the sustained recommendation came from the forge-verify rows; every gated dropday arm, the CLI turbo-default resolution for this public id, and the installed-app receipt (55.5 tok/s, draft 1.0) all served turbo. Served behavior unchanged: under turbo the hidden profile-scoped contract fell back to draft=target sampler (1.0/0.95/20), identical to the recommended_draft_sampler stamp now visible.",
|
| 92 |
"source_format": "bf16_native",
|
| 93 |
"source_repo": "/Users/youssof/.mtplx/models/Qwen--Qwen3.8-27B",
|
| 94 |
-
"source_sha": null
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 95 |
},
|
| 96 |
"mtp_contract": {
|
| 97 |
"base_hidden_variant": "post_norm",
|
|
@@ -103,8 +114,8 @@
|
|
| 103 |
},
|
| 104 |
"mtp_depth_default": 3,
|
| 105 |
"mtp_depth_max": 3,
|
| 106 |
-
"mtp_sidecar": "
|
| 107 |
-
"mtplx_version": "2.
|
| 108 |
"precision_policy": {
|
| 109 |
"intended_default_for": [
|
| 110 |
"m1",
|
|
@@ -144,14 +155,18 @@
|
|
| 144 |
"top_p": 0.95
|
| 145 |
},
|
| 146 |
"speed_evidence": {
|
|
|
|
|
|
|
|
|
|
|
|
|
| 147 |
"acceptance_by_depth": [
|
| 148 |
-
0.
|
| 149 |
-
0.
|
| 150 |
-
0.
|
| 151 |
],
|
| 152 |
-
"
|
| 153 |
-
|
| 154 |
-
|
| 155 |
"forge_verify_rows": [
|
| 156 |
{
|
| 157 |
"acceptance_by_position": [],
|
|
@@ -163,12 +178,12 @@
|
|
| 163 |
"hit_token_budget_count": 0,
|
| 164 |
"multiplier_vs_ar": 1.0,
|
| 165 |
"quality_passed": true,
|
| 166 |
-
"tok_s":
|
| 167 |
-
"verify_time_s":
|
| 168 |
},
|
| 169 |
{
|
| 170 |
"acceptance_by_position": [
|
| 171 |
-
0.
|
| 172 |
],
|
| 173 |
"depth": 1,
|
| 174 |
"finish_reasons": {
|
|
@@ -176,15 +191,15 @@
|
|
| 176 |
},
|
| 177 |
"hit_token_budget": false,
|
| 178 |
"hit_token_budget_count": 0,
|
| 179 |
-
"multiplier_vs_ar": 1.
|
| 180 |
"quality_passed": true,
|
| 181 |
-
"tok_s":
|
| 182 |
-
"verify_time_s":
|
| 183 |
},
|
| 184 |
{
|
| 185 |
"acceptance_by_position": [
|
| 186 |
-
0.
|
| 187 |
-
0.
|
| 188 |
],
|
| 189 |
"depth": 2,
|
| 190 |
"finish_reasons": {
|
|
@@ -192,16 +207,16 @@
|
|
| 192 |
},
|
| 193 |
"hit_token_budget": false,
|
| 194 |
"hit_token_budget_count": 0,
|
| 195 |
-
"multiplier_vs_ar": 1.
|
| 196 |
"quality_passed": true,
|
| 197 |
-
"tok_s":
|
| 198 |
-
"verify_time_s":
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"acceptance_by_position": [
|
| 202 |
-
0.
|
| 203 |
-
0.
|
| 204 |
-
0.
|
| 205 |
],
|
| 206 |
"depth": 3,
|
| 207 |
"finish_reasons": {
|
|
@@ -209,26 +224,23 @@
|
|
| 209 |
},
|
| 210 |
"hit_token_budget": false,
|
| 211 |
"hit_token_budget_count": 0,
|
| 212 |
-
"multiplier_vs_ar": 2.
|
| 213 |
"quality_passed": true,
|
| 214 |
-
"tok_s":
|
| 215 |
-
"verify_time_s":
|
| 216 |
}
|
| 217 |
],
|
| 218 |
-
"greedy_diagnostic": {
|
| 219 |
-
"tok_s": 20.991412027515594
|
| 220 |
-
},
|
| 221 |
"quality_rejected": [],
|
| 222 |
-
"
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
"
|
| 226 |
},
|
| 227 |
"verified_on": {
|
|
|
|
| 228 |
"hardware": "macOS-26.3.1-arm64-arm-64bit-Mach-O",
|
| 229 |
"machine_arch": "arm64",
|
| 230 |
"macos": "26.3.1",
|
| 231 |
-
"model": "Qwen3.8-27B-MTPLX-Optimized-Speed"
|
| 232 |
-
"timestamp": "2026-08-14T11:28:40-07:00"
|
| 233 |
}
|
| 234 |
}
|
|
|
|
| 91 |
"recommended_profile_note": "Restamped sustained->turbo 2026-08-14 23:50 (mirrors the OQ restamp): the sustained recommendation came from the forge-verify rows; every gated dropday arm, the CLI turbo-default resolution for this public id, and the installed-app receipt (55.5 tok/s, draft 1.0) all served turbo. Served behavior unchanged: under turbo the hidden profile-scoped contract fell back to draft=target sampler (1.0/0.95/20), identical to the recommended_draft_sampler stamp now visible.",
|
| 92 |
"source_format": "bf16_native",
|
| 93 |
"source_repo": "/Users/youssof/.mtplx/models/Qwen--Qwen3.8-27B",
|
| 94 |
+
"source_sha": null,
|
| 95 |
+
"head_quantization": {
|
| 96 |
+
"bits": 4,
|
| 97 |
+
"group_size": 64,
|
| 98 |
+
"mode": "affine",
|
| 99 |
+
"policy": "all",
|
| 100 |
+
"quantized_at": "2026-08-20T04:49:57Z",
|
| 101 |
+
"source_sidecar_bytes": 849400388,
|
| 102 |
+
"quantized_sidecar_bytes": 238934226,
|
| 103 |
+
"tool": "scripts/build_qwen38_q4head_sidecar.py",
|
| 104 |
+
"note": "Structural head quantization of the released sidecar; no calibration, no training. Trunk weights unchanged."
|
| 105 |
+
}
|
| 106 |
},
|
| 107 |
"mtp_contract": {
|
| 108 |
"base_hidden_variant": "post_norm",
|
|
|
|
| 114 |
},
|
| 115 |
"mtp_depth_default": 3,
|
| 116 |
"mtp_depth_max": 3,
|
| 117 |
+
"mtp_sidecar": "int4-g64-prequantized",
|
| 118 |
+
"mtplx_version": "2.9.0",
|
| 119 |
"precision_policy": {
|
| 120 |
"intended_default_for": [
|
| 121 |
"m1",
|
|
|
|
| 155 |
"top_p": 0.95
|
| 156 |
},
|
| 157 |
"speed_evidence": {
|
| 158 |
+
"depth": 3,
|
| 159 |
+
"tok_s": [
|
| 160 |
+
45.38375130593394
|
| 161 |
+
],
|
| 162 |
"acceptance_by_depth": [
|
| 163 |
+
0.9384057971014492,
|
| 164 |
+
0.855072463768116,
|
| 165 |
+
0.7427536231884058
|
| 166 |
],
|
| 167 |
+
"greedy_diagnostic": {
|
| 168 |
+
"tok_s": 19.60451265560769
|
| 169 |
+
},
|
| 170 |
"forge_verify_rows": [
|
| 171 |
{
|
| 172 |
"acceptance_by_position": [],
|
|
|
|
| 178 |
"hit_token_budget_count": 0,
|
| 179 |
"multiplier_vs_ar": 1.0,
|
| 180 |
"quality_passed": true,
|
| 181 |
+
"tok_s": 19.60451265560769,
|
| 182 |
+
"verify_time_s": 30.370817083050497
|
| 183 |
},
|
| 184 |
{
|
| 185 |
"acceptance_by_position": [
|
| 186 |
+
0.9712460063897763
|
| 187 |
],
|
| 188 |
"depth": 1,
|
| 189 |
"finish_reasons": {
|
|
|
|
| 191 |
},
|
| 192 |
"hit_token_budget": false,
|
| 193 |
"hit_token_budget_count": 0,
|
| 194 |
+
"multiplier_vs_ar": 1.5051947774660717,
|
| 195 |
"quality_passed": true,
|
| 196 |
+
"tok_s": 29.5086100639882,
|
| 197 |
+
"verify_time_s": 20.34781050006859
|
| 198 |
},
|
| 199 |
{
|
| 200 |
"acceptance_by_position": [
|
| 201 |
+
0.9889624724061811,
|
| 202 |
+
0.9448123620309051
|
| 203 |
],
|
| 204 |
"depth": 2,
|
| 205 |
"finish_reasons": {
|
|
|
|
| 207 |
},
|
| 208 |
"hit_token_budget": false,
|
| 209 |
"hit_token_budget_count": 0,
|
| 210 |
+
"multiplier_vs_ar": 1.976632257031939,
|
| 211 |
"quality_passed": true,
|
| 212 |
+
"tok_s": 38.75091209846504,
|
| 213 |
+
"verify_time_s": 34.64071964554023
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"acceptance_by_position": [
|
| 217 |
+
0.9384057971014492,
|
| 218 |
+
0.855072463768116,
|
| 219 |
+
0.7427536231884058
|
| 220 |
],
|
| 221 |
"depth": 3,
|
| 222 |
"finish_reasons": {
|
|
|
|
| 224 |
},
|
| 225 |
"hit_token_budget": false,
|
| 226 |
"hit_token_budget_count": 0,
|
| 227 |
+
"multiplier_vs_ar": 2.3149645238925305,
|
| 228 |
"quality_passed": true,
|
| 229 |
+
"tok_s": 45.38375130593394,
|
| 230 |
+
"verify_time_s": 21.260114512755536
|
| 231 |
}
|
| 232 |
],
|
|
|
|
|
|
|
|
|
|
| 233 |
"quality_rejected": [],
|
| 234 |
+
"acceptance_collapsed": [],
|
| 235 |
+
"failure_reasons": [],
|
| 236 |
+
"verdict": "mtp_depth_wins",
|
| 237 |
+
"artifact_fingerprint": "sha256:069c2de291fd15b130383119b13f60c45e0f78481a180cf51887ae87c3986b12"
|
| 238 |
},
|
| 239 |
"verified_on": {
|
| 240 |
+
"timestamp": "2026-08-19T23:10:32-07:00",
|
| 241 |
"hardware": "macOS-26.3.1-arm64-arm-64bit-Mach-O",
|
| 242 |
"machine_arch": "arm64",
|
| 243 |
"macos": "26.3.1",
|
| 244 |
+
"model": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16"
|
|
|
|
| 245 |
}
|
| 246 |
}
|