Instructions to use xunlinkx/Spark-X2.5-4B-oQ4e-fp16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use xunlinkx/Spark-X2.5-4B-oQ4e-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("xunlinkx/Spark-X2.5-4B-oQ4e-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

How to use xunlinkx/Spark-X2.5-4B-oQ4e-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 "xunlinkx/Spark-X2.5-4B-oQ4e-fp16"

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": "xunlinkx/Spark-X2.5-4B-oQ4e-fp16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use xunlinkx/Spark-X2.5-4B-oQ4e-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 "xunlinkx/Spark-X2.5-4B-oQ4e-fp16"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "xunlinkx/Spark-X2.5-4B-oQ4e-fp16"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "xunlinkx/Spark-X2.5-4B-oQ4e-fp16",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use xunlinkx/Spark-X2.5-4B-oQ4e-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 "xunlinkx/Spark-X2.5-4B-oQ4e-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 xunlinkx/Spark-X2.5-4B-oQ4e-fp16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use xunlinkx/Spark-X2.5-4B-oQ4e-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 "xunlinkx/Spark-X2.5-4B-oQ4e-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 "xunlinkx/Spark-X2.5-4B-oQ4e-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"

Spark-X2.5-4B — MLX oQ4e (fp16 source)

Model Details

Model Description

oQ4e-quantized MLX build (created by xunlinkx with the omlx pipeline) of XHToken/Spark-X2.5-4B for first-class, all-local serving on Apple Silicon via omlx (the MLX OpenAI-compatible server) and mlx-lm.

Attribution chain: XHToken/Spark-X2.5-4B (official base, Apache-2.0) → xunlinkx's oQ4e quantization (this repo).

This is not a new model — it is a quantization of the existing checkpoint, produced with omlx's quantize_oq_streaming(..., oq_level=4, enhanced=True) pipeline (128×512-block imatrix calibration):

Why we made it: run a full 4B Spark-X2.5-class model locally in ~2.5 GB of memory at usable decode rates (≈47–50 tok/s on an M3 Pro), with the attention gates and projection heads kept at higher precision so reasoning quality survives 4-bit compression. The -fp16 suffix records that the quantizer consumed full-precision float16 source weights (best-fidelity input), as opposed to bf16 or already-quantized sources.
Quantization: method oQ4e (omlx enhanced streaming imatrix calibration), mixed-bit per-layer layout embedded in config.json (quantization / quantization_config) so the exact layout is restored automatically on load.
Developed by: xunlinkx (oQ4e quantization + MLX build, using omlx)
Base model by: XHToken (Spark-X2.5-4B)
Shared by: xunlinkx
Model type: spark2_5 text-generation decoder (MLX quantized, model_type: spark2_5)
Language(s) (NLP): multilingual (base model; Spark-X2.5 pretraining, including English/Chinese coverage)
License: Apache 2.0 (inherits the base checkpoint license)

Compatibility — Spark support required (patch it yourself until mainline)

model_type: spark2_5 loads only where native support exists. Native support is still converging (as of 2026-09-15): mlx-lm has not merged #1891, and omlx main carries a vendored patch (#3606) that no tagged release includes yet. Until then:

mlx-lm users: checkout ml-explore/mlx-lm#1891 (gh pr checkout 1891 inside a mlx-lm clone, or pip install "mlx-lm @ git+https://github.com/xunlinkx/mlx-lm@spark2_5-native"; ab1806e-era alternative for omlx-scheduler compatibility: .../xunlinkx/mlx-lm@spark2_5-backport).
omlx users: spark support is on main since jundot/omlx#3606 (2026-09-14, vendored patch, self-disabling once the native module is present) — but no tagged release includes it yet (latest stable v0.6.4 and the 0.7.0.dev pre-releases all predate it). Run omlx from source main, or apply jundot/omlx#3671 for the oQ rules + mlx-lm pin (gh pr checkout 3671).

Loading this checkpoint with an mlx-lm that predates spark2_5 support fails with an unsupported-model-type error. No other requirements: standard MLX safetensors + config.json, no custom code needed once support is present.

Quantization detail (exact, from `config.json`)

5.178 weight-bits per parameter average (group_size 64, affine). Embeddings and norms stay fp16; per-group scales add ~0.25 bpw on disk.
Layout: 4-bit base; sigmoid head gates (self_attn.g_proj) kept at 5–8 bit on all 36 layers, fused q_k_v_proj / down_proj lifted to 5–6 bit on sensitive layers, out_proj at 4–5 bit, MLP gate/up at 4-bit.
Sibling levels in this family: oQ5e (6.098 bpw), oQ6e (7.013 bpw), oQ8e (8.653 bpw) — see xunlinkx/Spark-X2.5-4B-oQ{e}-fp16.

Sampling (official, per the XHToken model card)

temperature 1.0, top_p 0.95, top_k −1 (disabled); thinking mode default. (Early third-party guides suggesting 0.6–0.7 do not match the official card.)

Measured performance (Apple M3 Pro, 36 GB)

Smoke/serve test: 50.05 tok/s generation, 2.49 GB peak.
mlx_lm.generate cake prompt (1024 tokens): 47.35 tok/s, 2.53 GB peak.

Usage

mlx_lm.generate --model xunlinkx/Spark-X2.5-4B-oQ4e-fp16 \
  -p "The secret to baking a good cake is" -m 1024 \
  --temp 1.0 --top-p 0.95

omlx serve --host 127.0.0.1 --port 8005
# model id: Spark-X2.5-4B-oQ4e-fp16 (OpenAI-compatible /v1/chat/completions)

License

Apache-2.0, inherited from XHToken/Spark-X2.5-4B.

Downloads last month: 140

Safetensors

Model size

4B params

Tensor type

U32

F16

MLX

Hardware compatibility

4-bit

Model tree for xunlinkx/Spark-X2.5-4B-oQ4e-fp16

Base model

XHToken/Spark-X2.5-4B-Base

Finetuned

XHToken/Spark-X2.5-4B

Quantized

(37)

this model

Spark-X2.5-4B — MLX oQ4e (fp16 source)

Model Details

Model Description

Compatibility — Spark support required (patch it yourself until mainline)

Quantization detail (exact, from config.json)

Sampling (official, per the XHToken model card)

Measured performance (Apple M3 Pro, 36 GB)

Usage

License

Model tree for xunlinkx/Spark-X2.5-4B-oQ4e-fp16

Quantization detail (exact, from `config.json`)