Instructions to use philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-MLX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-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("philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-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 philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-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 "philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-MLX"

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": "philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-MLX"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-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 "philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-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 philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-MLX

Run Hermes

hermes

OpenClaw new

How to use philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-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 "philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-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 "philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-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"

MLX LM

How to use philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-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 "philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-MLX"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-MLX"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-MLX",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

GLM-5.2-Demolition-q4a4-soul-MLX / AGENTIC.md

philipjohnbasile

Upload AGENTIC.md with huggingface_hub

6961b99 verified 14 days ago

preview code

Raw

History Blame

3.42 kB

	# Run this as your local agentic coder

	The pitch: not the highest SWE-bench — the most reliable. The local agentic coder that can't break:
	zero malformed tool-calls (grammar-enforced, structurally impossible), compiler-steered every line,
	fabrication-proof `done` (re-runs the real tests — can't fake a pass), and elite across the whole stack
	(design · math · security · science), not just code. Reliability is the moat raw capability can't buy.

	## Serve it (OpenAI-compatible, on your Mac)
	```bash
	# raise the GPU ceiling once (or long agentic runs OOM):
	sudo sysctl iogpu.wired_limit_mb=122000
	# serve base + the core soul (OpenAI-compatible on :8080):
	GLM_STREAM_EVAL=0 python -m mlx_lm.server --model models/GLM-5.2-q3a4-v4 \
	--adapter-path adapters-soul2 --port 8080
	# optional: concise-thinking proxy (caps reasoning tokens -> faster, cleaner tool turns):
	python scripts/08_think_proxy.py --port 8081 --upstream http://localhost:8080
	```

	## Point your agent at it
	Anything that speaks OpenAI-compatible drops in against `http://localhost:8080/v1` (or `:8081` for the
	think-proxy), any model name:

	\| Agent \| Setup \|
	\|---\|---\|
	\| Cline (VS Code) \| Provider: OpenAI Compatible · Base URL `http://localhost:8080/v1` · any model id \|
	\| Aider \| `aider --openai-api-base http://localhost:8080/v1 --openai-api-key x --model glm-demolition` \|
	\| OpenCode \| add an OpenAI-compatible provider pointing at `:8080/v1` \|
	\| Cursor \| Settings → Models → override OpenAI Base URL to `:8080/v1` \|
	\| Claude Code \| needs an OpenAI→Anthropic shim (or `rapid-mlx`'s `ANTHROPIC_BASE_URL` trick) — Claude Code speaks the Anthropic API, not OpenAI \|

	Recommended settings: `temperature 0.6`, `top_p 0.95` for coding; `enable_thinking` on for hard problems,
	off (or the think-proxy budget) for fast tool turns.

	## Why it doesn't break (the reliability stack)
	- Grammar-constrained tool-JSON — invalid tokens get zero probability at each step, so a malformed call is
	structurally impossible (vs the field's best: "fewer malformed"). Speaks the 2026 strict-schema + MCP conventions.
	- Verified / compiler-steered decoding — a line that adds a type error is backtracked as it's written.
	- Fabrication-proof `done` — the agent re-runs the original tests before claiming success; it can't hallucinate a pass.
	- Integrity layer — test-tamper guard, 16-provider secret-scan, scope enforcement, slopsquat guard.
	- 51-tool ReAct agent — trajectory compaction + stall detection for long-horizon runs; the verifier mesh checks every output against its real tool.

	## Honest limits (and what's queued to fix them)
	- ~11–14 tok/s decode — the memory trade for a local 743B-class model. Queued: NVFP4 + M5 Neural Accelerators
	(~2× decode on M5), throughput batching for concurrent requests (2.6× at B=8 today).
	- Long single generations can degenerate at 3-bit (Computation Collapse). Queued: saliency-dynamic quant
	(#59 — early/late experts at 4-bit) + a serve-layer auto-recovery that re-structures broken tool-output.
	- Needs a 128 GB Mac — premium tier for now; a 64 GB sibling is deferred (depth before breadth).

	The trade we made on purpose: less raw benchmark, in exchange for reliability + breadth + fully local + verify-everything — the things that actually decide whether an agent finishes the job.

	# Run this as your local agentic coder

	The pitch: not the highest SWE-bench — the most reliable. The local agentic coder that can't break:
	zero malformed tool-calls (grammar-enforced, structurally impossible), compiler-steered every line,
	fabrication-proof `done` (re-runs the real tests — can't fake a pass), and elite across the whole stack
	(design · math · security · science), not just code. Reliability is the moat raw capability can't buy.

	## Serve it (OpenAI-compatible, on your Mac)
	```bash
	# raise the GPU ceiling once (or long agentic runs OOM):
	sudo sysctl iogpu.wired_limit_mb=122000
	# serve base + the core soul (OpenAI-compatible on :8080):
	GLM_STREAM_EVAL=0 python -m mlx_lm.server --model models/GLM-5.2-q3a4-v4 \
	--adapter-path adapters-soul2 --port 8080
	# optional: concise-thinking proxy (caps reasoning tokens -> faster, cleaner tool turns):
	python scripts/08_think_proxy.py --port 8081 --upstream http://localhost:8080
	```

	## Point your agent at it
	Anything that speaks OpenAI-compatible drops in against `http://localhost:8080/v1` (or `:8081` for the
	think-proxy), any model name:

	\| Agent \| Setup \|
	\|---\|---\|
	\| Cline (VS Code) \| Provider: OpenAI Compatible · Base URL `http://localhost:8080/v1` · any model id \|
	\| Aider \| `aider --openai-api-base http://localhost:8080/v1 --openai-api-key x --model glm-demolition` \|
	\| OpenCode \| add an OpenAI-compatible provider pointing at `:8080/v1` \|
	\| Cursor \| Settings → Models → override OpenAI Base URL to `:8080/v1` \|
	\| Claude Code \| needs an OpenAI→Anthropic shim (or `rapid-mlx`'s `ANTHROPIC_BASE_URL` trick) — Claude Code speaks the Anthropic API, not OpenAI \|

	Recommended settings: `temperature 0.6`, `top_p 0.95` for coding; `enable_thinking` on for hard problems,
	off (or the think-proxy budget) for fast tool turns.

	## Why it doesn't break (the reliability stack)
	- Grammar-constrained tool-JSON — invalid tokens get zero probability at each step, so a malformed call is
	structurally impossible (vs the field's best: "fewer malformed"). Speaks the 2026 strict-schema + MCP conventions.
	- Verified / compiler-steered decoding — a line that adds a type error is backtracked as it's written.
	- Fabrication-proof `done` — the agent re-runs the original tests before claiming success; it can't hallucinate a pass.
	- Integrity layer — test-tamper guard, 16-provider secret-scan, scope enforcement, slopsquat guard.
	- 51-tool ReAct agent — trajectory compaction + stall detection for long-horizon runs; the verifier mesh checks every output against its real tool.

	## Honest limits (and what's queued to fix them)
	- ~11–14 tok/s decode — the memory trade for a local 743B-class model. Queued: NVFP4 + M5 Neural Accelerators
	(~2× decode on M5), throughput batching for concurrent requests (2.6× at B=8 today).
	- Long single generations can degenerate at 3-bit (Computation Collapse). Queued: saliency-dynamic quant
	(#59 — early/late experts at 4-bit) + a serve-layer auto-recovery that re-structures broken tool-output.
	- Needs a 128 GB Mac — premium tier for now; a 64 GB sibling is deferred (depth before breadth).

	The trade we made on purpose: less raw benchmark, in exchange for reliability + breadth + fully local + verify-everything — the things that actually decide whether an agent finishes the job.