Instructions to use ukisai/Swift-1.5-4bit-MLX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ukisai/Swift-1.5-4bit-MLX 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("ukisai/Swift-1.5-4bit-MLX")

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 ukisai/Swift-1.5-4bit-MLX with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "ukisai/Swift-1.5-4bit-MLX"

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": "ukisai/Swift-1.5-4bit-MLX"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use ukisai/Swift-1.5-4bit-MLX with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "ukisai/Swift-1.5-4bit-MLX"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "ukisai/Swift-1.5-4bit-MLX"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "ukisai/Swift-1.5-4bit-MLX",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use ukisai/Swift-1.5-4bit-MLX 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 "ukisai/Swift-1.5-4bit-MLX"

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 ukisai/Swift-1.5-4bit-MLX

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use ukisai/Swift-1.5-4bit-MLX with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "ukisai/Swift-1.5-4bit-MLX"

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 "ukisai/Swift-1.5-4bit-MLX" \
  --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"

Swift-1.5-4bit-MLX / compatibility /architecture-compatibility-report.md

ukisai

initial release

9fd3d5f 13 days ago

preview code

Raw History Blame

2.47 kB

Swift 1.5 MLX compatibility: validated complete checkpoint

Source: ukisai/Swift-1.5-Qwen3.8-27b at 00ccd14e006897d28cb0ed5bf26390e60d274251. All 18 BF16 shards and original runtime assets passed full SHA256 verification. Source: 55563006776 weight bytes, 1199 BF16 tensors.

The official Qwen loader dropped 333 vision and 15 MTP tensors. The isolated patch adds mlx_lm/models/qwen3_5_full.py with a real vision encoder and explicit MTP module, plus strict dispatch/index checks and config/asset preservation in mlx_lm/utils.py. tests/test_qwen3_5_full.py exercises mapping failures, Transformers numerical agreement, cache behavior, native nonquantized conversion, and exact-target quantized saving/reloading.

All 16 nonquantized architecture tests and the additional fixed affine test passed on Linux CPU. The real source loaded strictly with 851 text, 333 vision and 15 MTP tensors. There are zero ignored or unexplained tensors. The native quant has 2379 saved tensors because quantized weights have scales and biases. All 609 BF16 remainder tensors equal the source values.

Text generation passed. Vision encoder execution and an MTP step using real text hidden states passed. Image/video text integration and speculative generation are not implemented. The original tokenizer, chat template, context, processor, untied output head/shared MTP embeddings, norms and gating configuration are retained.

CPU runtime checks promote only in-memory floating values to FP32. The original Linux BF16 QMM kernel produced 256 when summing 8192 exact ones; FP32 returned 8192. This reproducible backend issue and the runtime workaround are recorded in cpu-quantized-matmul-diagnostic.json and ../USAGE.md. Stored weights were not changed.

Apple Silicon Metal checks passed for all 2379 native parameter headers and real packed Q4 samples from text, vision and MTP. Both native BF16 Metal and FP32 Metal executions matched their references. Only small actual samples were evaluated on the 16 GiB Mac; no full 27B Mac generation is claimed.

The original source was read without modification. The source tensor layout transposes are recorded individually in quant-tensor-mapping-manifest.json. No custom quantization algorithm or alternate quantization configuration was used.

Patch SHA256: f6f1d0bdafa45863bfbf93dac0398c481c993ea04fdf38b9bae98c643f89eaec. Apply to official MLX-LM commit c69d1288440a0dc4e6401fc417098b07598dccd5; see ../USAGE.md.