Instructions to use lancejames221b/razorstrike-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lancejames221b/razorstrike-v1 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("lancejames221b/razorstrike-v1")
config = load_config("lancejames221b/razorstrike-v1")

# 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

How to use lancejames221b/razorstrike-v1 with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "lancejames221b/razorstrike-v1"

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": "lancejames221b/razorstrike-v1"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use lancejames221b/razorstrike-v1 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 "lancejames221b/razorstrike-v1"

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 lancejames221b/razorstrike-v1

Run Hermes

hermes

OpenClaw new

How to use lancejames221b/razorstrike-v1 with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "lancejames221b/razorstrike-v1"

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 "lancejames221b/razorstrike-v1" \
  --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"

RazorStrike-v1

⚠️ Provenance change (2026-07-21): this repo's weights were replaced. The current build is a 4-bit MLX quantization of nightmedia/Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx with a Qwen3.6 vision tower spliced back in — it is a different lineage than the DARE-TIES huihui-ai/SIQ-1 merge previously hosted here. This is a stopgap release for now, until RazorStrike-v2 (a from-scratch multi-domain LoRA fine-tune, in training) ships. The prior build's model card and investigation manifest are preserved as PRIOR_BUILD_README.md / PRIOR_BUILD_MANIFEST.json in this repo for reference — that work is not lost, just superseded as the active release.

⚠️ Known issue: repetition/looping in extended reasoning traces, inherited from this architecture's reasoning-donor lineage. repetition_penalty: 1.1 (baked into generation_config.json) mitigates but doesn't fully eliminate it. RazorStrike-v2 trains directly on anti-loop data to address this at the source.

What this build actually is

Base merge (not built by this repo — credit to nightmedia): Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx, itself a merge of nightmedia/Qwen3.6-35B-A3B-MTP-Holo3-Qwopus-BF16 (which traces to Hcompany/Holo3-35B-A3B, samuelcardillo/Qwopus-MoE-35B-A3B, and llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved) with Qwen/Qwen-AgentWorld-35B-A3B.
Vision tower restored: the source merge's public release was text-only (no vision tensors, despite image-text-to-text metadata carried over from the base template). This repo splices the original Qwen3.6-35B-A3B vision tower (ViT blocks + merger/connector, ~333 tensors) back onto the merged language model, verified functional end-to-end (object/color identification, OCR on rendered text) before and after quantization.
Quantized: dequantized the source's mixed-precision weights to bf16, added the vision tower, then requantized the full vision+language model to clean uniform 4-bit (group-size 64, affine) using mlx_vlm (vision-aware — mlx_lm's text-only converter would silently mishandle or drop the vision tensors).

Verification

Vision: tested on synthetic images (shape/color identification, rendered-text OCR) at both bf16 (pre-quant) and 4-bit (post-quant) — correct on both, no quantization-induced degradation observed in these tests.
Text: loads and generates coherently through both the standalone mlx_vlm CLI and the deployed LM Studio OpenAI-compatible API.
Not independently re-run: the original coherence/refusal battery documented for the prior DARE-TIES build (see PRIOR_BUILD_MANIFEST.json) was validated against that build's weights, not this one. This build has not been put through that same battery — treat behavior as inherited-but-unverified from the nightmedia source pending further testing.

Recommended inference settings

temperature: 0.6
top_p: 0.95
top_k: 20
min_p: 0.0
repetition_penalty: 1.1

Baked into generation_config.json. Context: 262,144 tokens (inherited from the Qwen3.6-35B-A3B lineage).

Usage

from mlx_vlm import load, generate
model, processor = load("lancejames221b/razorstrike-v1")
# see mlx_vlm docs for image + chat-template usage

Also loads directly in LM Studio (shows as a VLM, image upload supported). For a GGUF/llama.cpp build of this exact lineage, see lancejames221b/razorstrike-v1-GGUF. The unrelated, superseded DARE-TIES lineage's full-precision master is preserved separately at lancejames221b/razorstrike-v1-bf16 (not the source for this repo).