Workflows/Run GLM-5.3 with Strix for Authorized Security Reviews

Agents / WORKFLOW

Run GLM-5.3 with Strix for Authorized Security Reviews

Choose GLM-5.3 Flash or full GLM-5.3, then run Strix with a self-hosted Linux cluster, rented RunPod GPUs, or the OpenRouter hosted API to review a disposable repository you are authorized to assess.

Verified sourceSource checked 10/14/2026

LOCALRENTED GPUHOSTED API

Freshness 82 out of 100. Repository age: 51 days old, Aug 25, 2026. Measures age, not quality.

Popularity 85 out of 100. 1,803,969 downloads in 30 days; 2,223 lifetime likes. Checked Oct 15, 2026. Measures adoption, not quality.

Arena Open Source Agents: rank 6 of 18. Ranked model: GLM 5.3 (Max). Original model ranking, not a benchmark of this package. Export dated Oct 14, 2026. Source checked Oct 15, 2026. Human preference ranking, not hardware suitability.

Category rankings & sources

Each position belongs to a separate Arena category. These are human preference rankings—not hardware recommendations. We do not combine the categories into one score.

Agents#6of 18 open models

GLM 5.3 (Max)zaiLicense: MIT

Original model ranking. This does not benchmark the selected quantization or package.

Export dated Oct 14, 2026Checked: Oct 15, 2026Overall board: #24

View official Agents board

Web coding#6of 55 open models

glm-5.3-maxzaiLicense: MIT

Original model ranking. This does not benchmark the selected quantization or package.

Export dated Oct 13, 2026Checked: Oct 15, 2026Overall board: #23

View official Web coding board

Text#3of 225 open models

glm-5.3-maxzaiLicense: MIT

Original model ranking. This does not benchmark the selected quantization or package.

Export dated Oct 13, 2026Checked: Oct 15, 2026Overall board: #28

View official Text board

Rankings reflect the dated export; live Arena boards can change. Open-source positions use Arena’s open-model filter. The overall position also includes the other models on that category board. Availability of weights and license terms still need to be checked separately.

What do these scores mean?

Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.

Freshness · release age

A new release starts at 100. The score halves every 180 days: 100 × 2^(−age in days / 180). When the release date is unavailable, a repository date is a fallback—not proof of when the model was released. Missing dates stay unscored.

Check the date source

Popularity · adoption

Downloads over the past 30 days carry 80% of the score; lifetime likes carry 20%. Both use a logarithmic scale, capped at 10 million downloads and 100,000 likes. If likes are unavailable, downloads carry all the weight. Missing download statistics stay unscored.

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source

Scores as of Oct 15, 2026. Statistics checked Oct 15, 2026. Statistics are marked for refresh after 72 hours.Full scoring method

Before you begin

Fastest route: OpenRouter

Strix can use the hosted GLM model without a local GPU or weight download. The hardware tables below apply only if you choose to self-host on Linux or rent matching GPUs.

Jump to the setup choices

Hardware

Self-hosted native-FP8 requirements only; OpenRouter needs no local GPU. These are publisher deployment routes, not YouRunAI speed tests. Memory is per GPU, with extra room needed for runtime and context.

Machine	GLM-5.3 Flash · native FP8	GLM-5.3 · native FP8
NVIDIA · 1× RTX 5090 32 GB	Cannot hold ~306 GiB native FP8 weights; use the 4×GB200 route or a smaller checkpoint.	Cannot hold ~750 GB native FP8 weights; rent 8×H200.
NVIDIA · 4× GB200 192 GB each	Main publisher vLLM route; start at 16K context.	Check the separate full-model recipe and topology before running.
NVIDIA · 4× H200 141 GB each	Only the specialized text-only KV-offload route was validated here; this guide uses 4×GB200.	Insufficient for full native FP8; choose 8×H200.
NVIDIA · 8× H200 141 GB each	More memory than the main 4×GB200 route; verify supported configuration and cost.	Documented vLLM route; start at 128K context.
AMD · 4× MI355X 288 GB each	Documented ROCm route; verify current build.	Full-model route requires 8×MI325X or equivalent validated capacity.
AMD · 8× MI325X 256 GB each	More than Flash needs; check the AMD recipe.	Documented ROCm route; verify current build.
Intel GPU · 80 GB each	No validated native-FP8 Intel vLLM route; use a documented NVIDIA or AMD cluster.	No validated native-FP8 Intel vLLM route; use a documented NVIDIA or AMD cluster.
Apple Silicon · 512 GB unified memory	No documented native-FP8 Apple Silicon vLLM server route; use Mac as SSH client.	No documented native-FP8 Apple Silicon vLLM server route; use Mac as SSH client.
CPU only · 1 TB system RAM	No practical validated CPU route for this guide; use 4×GB200 or a smaller model.	No practical validated CPU route for this guide; rent 8×H200.

A 1M advertised context is an architectural limit, not a promise that it fits these starting configurations. Rented Pod availability and price must be checked live.

Context window

Start small enough to leave KV-cache headroom. Increase only after a successful tool-call test and checking free memory.

Machine	GLM-5.3 Flash · native FP8	GLM-5.3 · native FP8
NVIDIA · 4× GB200 192 GB each	Start at 16K in this guide and increase only after memory checks.	Check the separate full-model recipe.
NVIDIA · 4× H200 141 GB each	A specialized text-only offload configuration was validated at 16K; this guide does not implement that route.	Insufficient for full native FP8.
NVIDIA · 8× H200 141 GB each	Check the exact Flash configuration rather than assuming a 4×H200 default.	Start at 128K per the setup command; lower if memory is tight.
AMD · 4× MI355X 288 GB each	Start at 16K and check the ROCm recipe.	Use a validated larger cluster.
AMD · 8× MI325X 256 GB each	Start at 16K and check the ROCm recipe.	Start at 128K; lower if memory is tight.

These are conservative first-run choices, not measured throughput or maximum supported context.

OPENROUTER / RUNPOD / PRIVATE SERVERSOURCE-LINKED SETUP

Review authorized code and apps with Strix and GLM-5.3.

Start with OpenRouter + full GLM-5.3: no GPU or model download. Switch to Flash for a lighter hosted option, or serve either native FP8 model on RunPod or your own Linux GPU cluster. Strix can review an authorized repository, staging site, API, or both; verify every finding.

Choose a package Run its commands Check the result

BEFORE YOU BEGIN

Choose your package

Choose a hosted GLM model. Your selection changes the OpenRouter model ID in the steps below.

OpenRouter runs the selected model for you. No local GPU, weight download, or rented Pod is needed; usage is billed by OpenRouter. Choose either model package below to change its API route.

CURRENT SETUP

GLM-5.3 · official

Use z-ai/glm-5.3 through OpenRouter. This route needs no local GPU, weight download, or RunPod rental. OpenRouter bills API usage; review its current model availability, price, and privacy terms before submitting code.

Freshness 82 out of 100. Repository age: 51 days old, Aug 25, 2026. Measures age, not quality.

Popularity 85 out of 100. 1,803,969 downloads in 30 days; 2,223 lifetime likes. Checked Oct 15, 2026. Measures adoption, not quality.

Category rankings & sources

Each position belongs to a separate Arena category. These are human preference rankings—not hardware recommendations. We do not combine the categories into one score.

Agents#6of 18 open models

GLM 5.3 (Max)zaiLicense: MIT

Original model ranking. This does not benchmark the selected quantization or package.

Export dated Oct 14, 2026Checked: Oct 15, 2026Overall board: #24

View official Agents board

Web coding#6of 55 open models

glm-5.3-maxzaiLicense: MIT

Original model ranking. This does not benchmark the selected quantization or package.

Export dated Oct 13, 2026Checked: Oct 15, 2026Overall board: #23

View official Web coding board

Text#3of 225 open models

glm-5.3-maxzaiLicense: MIT

Original model ranking. This does not benchmark the selected quantization or package.

Export dated Oct 13, 2026Checked: Oct 15, 2026Overall board: #28

View official Text board

What do these scores mean?

Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.

Freshness · release age

Check the date source

Popularity · adoption

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source

Scores as of Oct 15, 2026. Statistics checked Oct 15, 2026. Statistics are marked for refresh after 72 hours.Full scoring method

View OpenRouter model

THE WORKFLOW

Connect GLM-5.3 through OpenRouter

The model runs through a hosted API; follow these steps on your own computer. The instructions below update to match your selection.

CHOOSE YOUR SETUPOpenRouter

Install Docker for your operating system

Strix uses a Docker sandbox. On Windows, install and start Docker Desktop with Linux containers; use PowerShell for the Windows commands below. Docker Desktop may use WSL 2 behind the scenes, but this guide does not require an Ubuntu terminal. On Mac, use Docker Desktop and Terminal. On Linux, use Docker Engine and Terminal. Continue only when Docker reports engine information.

Docker for Windows Docker for Mac Docker Engine for Linux

docker --version; docker info

Windows PowerShell: install Strix

Skip this step on Mac and Linux. Install Python 3.12 or newer if needed. Run these PowerShell commands to install pipx and the published Strix Windows package. Reopen PowerShell after ensurepath so strix is available, then check strix --version. WSL/Ubuntu is not required to run the Strix client.

Python on Windows Install pipx on Windows Strix Windows package

py -m pip install --user pipx
py -m pipx ensurepath
py -m pipx install strix-agent

Mac or Linux: install Strix

Skip this step on Windows. In Mac or Linux Terminal, run the official Bash installer after Docker works, then open a fresh Terminal and check strix --version. Keep the repository you will review in a disposable, secret-free copy.

Official Strix quickstart

curl -sSL https://strix.ai/install | bash

Choose z-ai/glm-5.3 on OpenRouter

Open the exact model listing and check current availability and usage price. Create an OpenRouter API key. This is hosted inference: code sent in prompts is processed by the selected API provider, unlike the self-hosted routes above.

Exact OpenRouter model Create an OpenRouter API key

Windows PowerShell: enter your API key privately

Skip this step on Mac and Linux. PowerShell asks for your OpenRouter key without displaying it, then sets LLM_API_KEY only in this PowerShell process for Strix. Do not paste a key into a file or into this guide. The environment variable holds plain text while Strix runs; clear it after the scan.

Strix OpenRouter setup

$secureKey = Read-Host "OpenRouter API key" -AsSecureString; $env:LLM_API_KEY = [System.Net.NetworkCredential]::new("", $secureKey).Password; Remove-Variable secureKey

Mac or Linux Bash: enter your API key privately

Skip this step on Windows. In a Bash Terminal, read your OpenRouter key without showing it on screen. On Mac, type bash first if your Terminal is using zsh. Keep the key out of files and command history.

Strix OpenRouter setup

read -rsp "OpenRouter API key: " LLM_API_KEY && export LLM_API_KEY && printf "\n"

Windows PowerShell: select the hosted model

Skip this step on Mac and Linux. In the same PowerShell window that holds your key, clear any earlier local-model URL and select OpenRouter's z-ai/glm-5.3. The last command prints only the model ID, not the key. Switching packages on this page changes this exact model ID.

Official Strix OpenRouter configuration

Remove-Item Env:LLM_API_BASE -ErrorAction SilentlyContinue; $env:STRIX_LLM="openrouter/z-ai/glm-5.3"; $env:STRIX_REASONING_EFFORT="high"; $env:STRIX_LLM

Mac or Linux Bash: select the hosted model

Skip this step on Windows. In the same Bash Terminal that holds your key, clear any earlier local-model URL and select OpenRouter's z-ai/glm-5.3. The command prints only the model ID, not the key.

Official Strix OpenRouter configuration Strix configuration

unset LLM_API_BASE; export STRIX_LLM="openrouter/z-ai/glm-5.3" STRIX_REASONING_EFFORT="high"; printf '%s\n' "$STRIX_LLM"

Run a first, budgeted review

Make a disposable, secret-free copy of a repository you own or have written permission to assess. From its parent directory, run a quick scan with a $10 maximum LLM-spend setting; change or omit that cap yourself. Strix can write to the mounted copy. This is a first-run check, not a complete assessment.

Strix targets, scan modes, and budget flag

strix -n -t ./authorized-copy --scan-mode quick --max-budget 10

Windows PowerShell: clear the temporary key

Skip this step on Mac and Linux. After reviewing the report and any file changes, remove the key from this PowerShell process. Check your OpenRouter usage page for the cost of the review.

OpenRouter key and usage settings

Remove-Item Env:LLM_API_KEY -ErrorAction SilentlyContinue

Mac or Linux Bash: clear the temporary key

Skip this step on Windows. After reviewing the report and any file changes, clear the key from this Terminal and check OpenRouter usage.

OpenRouter key and usage settings

unset LLM_API_KEY

MORE WAYS TO USE THIS SETUP

Choose the target you are allowed to test

Strix accepts code, a deployed app, an API specification, or multiple targets. Use a disposable copy for local code: Strix mounts it writable. Replace the example paths and domains with assets you own or have explicit permission to assess.

Local repository copy

Use this for a private repo or a first connectivity check. Remove secrets from the copy and review its file diff afterward.

Strix local target behavior

strix -n -t ./authorized-copy --scan-mode quick --max-budget 10

GitHub repository

Use a repository you own or are authorized to review. A remote repository is a different target from a local directory; use the exact URL.

Supported target types

strix -n -t https://github.com/OWNER/REPO --scan-mode quick --max-budget 10

Authorized staging site or API

Use only a hostname within the agreed scope. Start with quick mode. Do not target production without a clear testing window and permission.

Strix target syntax

strix -n -t https://staging.example.com --scan-mode quick --max-budget 10

Source plus deployed target

Give Strix both an authorized code copy and its staging site for a white-box review. Both target arguments belong to the same scan.

Strix multiple-target example

strix -n -t ./authorized-copy -t https://staging.example.com --scan-mode quick --max-budget 15

Private repository

Clone into a disposable directory using your usual Git authentication, remove secrets, and run the local-directory route above. Never embed credentials in the guide or command history.

Local directory handling

git clone YOUR_PRIVATE_REPOSITORY_URL ./authorized-copy

MORE WAYS TO USE THIS SETUP

Choose depth, cost, and focus

Start with quick mode to prove the setup. The optional --max-budget flag limits estimated LLM spend for one scan, not your YouRunAI site or OpenRouter account; provider billing may vary. Increase depth only after the first run completes.

Quick: prove the setup

Fastest starting point for one authorized code copy. A completed quick scan is a setup check and focused review, not proof the application is secure.

Strix scan modes

strix -n -t ./authorized-copy --scan-mode quick --max-budget 10

Standard: normal assessment

Once the provider, Docker sandbox, and tool calls work, use standard mode on a copy you can safely inspect afterward.

Strix scan modes

strix -n -t ./authorized-copy --scan-mode standard --max-budget 25

Deep: longer assessment

Use when you have time and an explicit scope. Deep scans can consume substantially more tokens. Inspect the earlier quick/standard reports first.

Strix budget behavior

strix -n -t ./authorized-copy --scan-mode deep --max-budget 50

Give the agent a precise scope

Focus on exploitable authentication, authorization, injection, file-handling, secrets, and dependency issues. Exclude destructive testing, denial of service, persistence, and any hostname outside the written authorization.

Strix custom instructions

strix -n -t ./authorized-copy --scan-mode standard --max-budget 25 --instruction "Review only this authorized copy. Prioritize exploitable auth, IDOR, injection, secrets, and unsafe file handling. Validate findings. No destructive actions or out-of-scope hosts."

MORE WAYS TO USE THIS SETUP

Tune reasoning and context

Strix uses long prompts and tool schemas. Hosted API context is managed by the provider; self-hosted context is constrained by available GPU memory. These are starting settings, not measured throughput.

Reasoning effort

Use high for a normal security review. Low can shorten a first connectivity check; max may increase analysis and cost if supported by the selected route.

Strix configuration settings

export STRIX_REASONING_EFFORT="high"

Flash on GB200: start at 16K

This guide follows the publisher’s four-GPU GB200 Flash route and begins at 16,384 tokens. Increase only after tool-call tests and memory checks. The separate 4×H200 text-only KV-offload validation uses special settings that are not implemented by this default path.

Flash deployment recipe and H200 offload note

Full GLM-5.3: start at 128K

This guide starts full GLM-5.3 at 131,072 tokens on 8×H200. Reduce the server context if KV-cache memory is exhausted. Treat 200K–400K and the architectural 1M as experimental targets until measured on your server.

Full-model deployment recipe

Tool calls are the real compatibility check

Before a self-hosted scan, test a chat request with a tool definition and confirm the response contains a structured tool_calls field. Plain text that looks like a tool call will not drive the Strix agent.

Strix local-model tool-call requirements

MORE WAYS TO USE THIS SETUP

Read the result and recover from setup failures

A healthy run advances beyond “Starting agent”, performs tool activity, and saves reports under strix_runs. Treat generated security findings as leads until you verify their evidence.

Check the final report

Open the run directory under strix_runs, confirm the target and scan mode, inspect cited source lines, and review the disposable repository diff. A budget stop is not a completed security review.

Strix quickstart and result location Budget stop behavior

Docker unavailable on macOS

First confirm Docker Desktop is running and docker info succeeds in a fresh Terminal. If Strix still cannot find the socket, the documented Docker Desktop socket path can be supplied to this Terminal for a retry.

Docker Desktop for Mac setup

export DOCKER_HOST="unix://$HOME/.docker/run/docker.sock"

Agent stalls on startup

Check provider logs and whether the model received the request. For a self-hosted endpoint, test /v1/models and a structured tool call. Check context-window and KV-cache errors before increasing GPU count or rerunning the scan.

Strix local-model troubleshooting

Flash sparse-MLA or plain-text tool-call error

Use the current GLM-5.3 Flash recipe and compatible vLLM/FlashInfer build. Keep --enable-auto-tool-choice with the glm47 tool and reasoning parsers. If the server emits tool syntax as text, repair its parser before running Strix.

Official Flash vLLM recipe Strix tool-call requirements

SUCCESS CHECK · GLM-5.3 · native FP8

A scoped Strix review with verifiable evidence

For self-hosted and RunPod routes, confirm the private /v1/models endpoint and structured tool_calls. For OpenRouter, confirm Strix authenticates with the selected hosted model. Every route should produce a report in strix_runs from the disposable authorized repository. Verify findings against source lines and check for unintended file changes.

Use this only on repositories and systems you own or have explicit authorization to assess. Strix can write to a mounted local target; start with a disposable copy without secrets. Self-hosted model serving requires a suitable Linux GPU cluster; Windows and Mac can be Strix clients. OpenRouter sends prompts to a hosted API and charges for usage. Publisher specifications are source-verified, not a YouRunAI hands-on benchmark.

Back to setup steps

When it doesn’t go to plan

vLLM cannot load the model or runs out of GPU memory.

Check that all GPUs and their memory are visible, that the exact package and compatible vLLM/ROCm/CUDA build match the official recipe, and that disk holds the weights. Reduce max-model-len first; Flash on H200 must not force FP8 KV cache. Do not try to squeeze the native FP8 checkpoints onto one consumer GPU.

Strix cannot call tools or the model emits tool syntax as text.

Confirm /v1/models returns the served repository ID, then verify a structured tool_calls response. Keep --enable-auto-tool-choice and --tool-call-parser glm47 from the official recipe. Set STRIX_LLM to openai/ followed by that exact model ID and LLM_API_BASE to the local /v1 endpoint.

The rented Pod is running but Strix cannot reach the model.

Check the Pod server is bound to 127.0.0.1:8000, use a Public IP Pod with full SSH and an active -L 8000:127.0.0.1:8000 tunnel, then retry /v1/models on your own computer. Do not expose raw vLLM HTTP to the internet.

Strix cannot connect to OpenRouter or says the API key is missing.

Confirm the selected OpenRouter model is available, the account has credits, and LLM_API_KEY is set in the same Terminal running Strix. STRIX_LLM must be openrouter/z-ai/glm-5.3 or openrouter/z-ai/glm-5.3-flash. Remove any stale LLM_API_BASE from an earlier local setup. Do not paste or print the key when troubleshooting.

Docker works in Terminal but Strix cannot find it on macOS.

Start Docker Desktop, open a fresh Terminal, and run docker info. If the CLI is not on PATH, use Docker Desktop Settings → Advanced to install CLI tools in the system path. If Strix still cannot access the socket, try DOCKER_HOST=unix://$HOME/.docker/run/docker.sock in that Terminal; do not make permanent shell changes until the temporary fix is confirmed.

The scan reports that the request exceeds the available context.

For self-hosting, check vLLM logs and available GPU/KV-cache memory. Increase --max-model-len only within measured capacity, or reduce task scope and use a shorter scan. The Flash 16K start is a server smoke test, not a guarantee that a full Strix run fits; the advertised 1M context is not a practical default.

A security finding lacks evidence or the local target changed unexpectedly.

Stop the run, inspect strix_runs and the disposable repository diff, and independently verify every claimed finding. Strix can write to a mounted local target; use an authorized throwaway copy without secrets and keep the task scope narrow.

REFERENCE LIBRARY

Sources and files

Original instructions, model files, and compatibility notes behind this setup.

12 SOURCES

01GLM-5.3 publisher model cardhuggingface.co 02GLM-5.3 Flash publisher model cardhuggingface.co 03Full-model vLLM deployment reciperecipes.vllm.ai 04Flash vLLM deployment reciperecipes.vllm.ai 05Strix quickstartdocs.strix.ai 06Strix local model configurationdocs.strix.ai 07Strix OpenRouter configurationdocs.strix.ai 08OpenRouter GLM-5.3openrouter.ai 09OpenRouter GLM-5.3 Flashopenrouter.ai 10Strix CLI and local target behaviordocs.strix.ai 11RunPod full SSH instructionsdocs.runpod.io 12vLLM endpoint securitydocs.vllm.ai

The model behind this workflow

GLM-5.3

Keep exploring

Text

SETUP GUIDE

Qwen3.5 · 9B private chat setup

Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.

Freshness 41 out of 100. Release age: 230 days old, Feb 27, 2026. Measures age, not quality.

Build a private coding assistant with Qwen3.8 27B

Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.

Freshness 76 out of 100. Repository age: 71 days old, Aug 5, 2026. Measures age, not quality.

Arena Open Source Web coding: rank 10 of 55. Ranked model: qwen3.8-27b. Original model ranking, not a benchmark of this package. Export dated Oct 13, 2026. Source checked Oct 15, 2026. Human preference ranking, not hardware suitability.+3

SOFTWAREOllama + Visual Studio Code + Cline + Python

Create images locally with FLUX.2 Klein

Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.

Freshness 35 out of 100. Release age: 273 days old, Jan 15, 2026. Measures age, not quality.

SOFTWAREComfyUI

LOCALRENTED GPU

Verified source