Connect a PC client to GLM-5.3 through an API. Hosted options use MagmaRouter’s Xploded or Abliteration.ai’s Large V2 derivatives. The exact dealignai FP8 route uses one eight-H200 RunPod server, persistent storage, a private SSH tunnel, and a supplied API check.
EstimatedSource checked 10/23/2026
RENTED GPU
Freshness 81 out of 100. Repository age: 54 days old, Aug 30, 2026. Measures age, not quality.81/ 100Freshness54 days oldRepository age · Aug 30, 2026
Popularity 48 out of 100. 2,698 downloads in 30 days; 158 lifetime likes. Checked Oct 23, 2026. Measures adoption, not quality.48/ 100Popularity2.7K downloads / 30d158 lifetime likes · Checked Oct 23, 2026
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.
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.
Scores as of Oct 23, 2026. Statistics checked Oct 23, 2026. Statistics are marked for refresh after 72 hours.Full scoring method →
Before you begin
Hardware
Your computer is the client. The exact FP8 model runs on the remote Linux GPU host; Xploded and Abliteration Large V2 run on their respective providers’ infrastructure.
Machine
Exact dealignai FP8
Hosted API alternatives
RunPod · 8× H200 141 GB on one Linux host
Documented checkpoint topology; start at 64K context and one sequence.
No GPU rental needed; use the provider API.
NVIDIA PC · one RTX 5090 32 GB
Cannot hold these FP8 weights; use this PC as an SSH/API client.
Client only; its GPU is not used for inference.
AMD PC · one GPU 24 GB VRAM
No exact AMD self-host route verified here; connect to the H200 server.
Client only; its GPU is not used for inference.
Intel PC · one GPU 16 GB VRAM
No exact Intel self-host route verified here; connect to the H200 server.
Client only; its GPU is not used for inference.
Apple Silicon Mac · 32 GB unified memory
Use macOS as the SSH/API client; no local FP8 server path is offered.
Client only; no local model download.
CPU-only PC · 16 GB system RAM
Use as the SSH/API client; no practical CPU model-server route is offered.
Client only; no local model download.
The model card specifies 8×H200. This page does not pool unrelated GPUs or claim measured latency, throughput, or an availability guarantee.
Context window
Context contains the prompt, history, tool schemas, and generated output. Keep room for all four instead of filling the window with input.
Machine
Exact dealignai FP8
Hosted API alternatives
RunPod · 8× H200 141 GB Linux server
Start at 65,536 total context; client output cap 8,192. Longer context uses more GPU memory.
Both providers advertise 1M; use 65,536 / 8,192 for the first small coding task.
The self-hosted settings are conservative starting choices. The hosted limit is provider-reported and does not prove equivalent weights or behavior.
PRIVATE MODEL API / HOSTED ALTERNATIVE / PC CLIENTSOURCE-LINKED SETUP
Run GLM-5.3 UNCENSORED through an API
Use an API from your own computer. For a quick hosted trial, choose MagmaRouter’s GLM-5.3 Xploded or Abliteration.ai’s abliterated-model-large-v2 below; both use a URL, key, and model ID with no GPU rental. To run the exact downloadable dealignai checkpoint, follow the eight-H200 RunPod route. These are different model builds. The prepared server setup and API checker reduce manual setup; this guide does not claim a completed GPU or hosted-generation test.
Choose a package Run its commands Check the result
01
BEFORE YOU BEGIN
The documented package
This setup uses the package documented for this task. Review its source and supported platforms before starting.
These requirements describe the GPU you rent. Use Windows, Mac or Linux as the browser client; your saved computer’s VRAM does not limit this setup.
One 8×H200 141 GB Linux server and at least 1.2 TB persistent storage. Start at 65,536 context tokens, one active sequence, and 8,192 output tokens. These are source-backed settings and capacity guidance, not a physical test or a proven minimum.
Freshness 81 out of 100. Repository age: 54 days old, Aug 30, 2026. Measures age, not quality.81/ 100Freshness54 days oldRepository age · Aug 30, 2026
Popularity 48 out of 100. 2,698 downloads in 30 days; 158 lifetime likes. Checked Oct 23, 2026. Measures adoption, not quality.48/ 100Popularity2.7K downloads / 30d158 lifetime likes · Checked Oct 23, 2026
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.
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.
Set up GLM-5.3 UNCENSORED · FP8 from your own computer
Run the app on your own computer; only model inference runs on the rented GPU. The instructions below update to match your selection.
CHOOSE YOUR SETUPRented GPU
The model runs on one rented Linux server containing eight H200 GPUs. Your Windows, macOS, or Linux computer runs the API client and optional VS Code/Cline. A local gaming GPU is not required. Source-checked instructions; no YouRunAI GPU execution is claimed.
01
Choose exact weights or the simpler hosted trial
Use this RunPod route when you want the downloadable dealignai checkpoint itself. For a quick first trial without GPU rental, choose the hosted MagmaRouter Xploded or Abliteration.ai Large V2 option below. Both providers supply their own GLM-5.3 derivatives, rather than hosted copies of the pinned dealignai revision. Keep the selected build’s identity in a comparison or video.
In RunPod, choose Secure Cloud, On-Demand, and one Pod with 8× H200 141 GB GPUs on the same host with a suitable interconnect. Eight separate Pods are not equivalent. Choose a Public IP and full SSH support. Set the container image to runpod/pytorch:1.2.0-cu1290-torch2130-ubuntu2204 from RunPod’s official image list. It includes the CUDA 12.9 development toolkit; choose a host driver supporting CUDA 12.9. The installer manages Python 3.12. Prefer 1 TB host RAM as a conservative allocation, not a measured minimum. On October 22, 2026, $5.29 per H200-hour gives $42.32/hour for eight GPUs, or $169.28 for four hours before storage. Confirm your actual quote and capacity before deploying.
Setting
Value
Container image
runpod/pytorch:1.2.0-cu1290-torch2130-ubuntu2204
CUDA compatibility
Image toolkit/nvcc 12.9; host NVIDIA driver must support CUDA 12.9
Before clicking Deploy, configure at least 1.2 TB of persistent storage at /workspace and a 100 GB container disk for the operating system and tools. For a network volume, create it first in a data center with the required GPU inventory and attach it when creating the Pod. The exact checkpoint is about 756 GB; runtime, cache, and temporary files need extra room. Do not put it only on the container disk. Storage is billed separately. A Pod volume is deleted with that Pod; a network volume has its own lifecycle. Confirm these choices, then deploy the Pod.
Setting
Value
Persistent mount
/workspace
Starting storage allocation
At least 1.2 TB; inspect actual free space before downloading
Add your public SSH key in RunPod and open the Pod’s Connect panel. Copy its “SSH over exposed TCP” command using the Pod’s public IP and mapped SSH port. Run that command in Windows PowerShell, macOS Terminal, or Linux Terminal on your own computer. Do not use the basic ssh.runpod.io proxy for this guide: the private API connection needs SSH port forwarding. You are now in the Pod’s Linux shell; the next command runs there.
In the Pod’s Linux Terminal, download the prepared setup and launch it in the background. The first launch installs pinned vLLM 0.32.0 for CUDA 12.9 and downloads the pinned checkpoint into /workspace; later launches reuse its environment and cache. Watch the saved log until server startup completes. Ctrl+C leaves the log view without stopping the model. The API listens only on 127.0.0.1:8000 inside the Pod, and the background server survives an SSH disconnect.
The prepared setup starts at 65,536 tokens of total context with one active sequence. Client output is capped at 8,192 tokens, leaving room for prompts, tool definitions, and prior turns. Low reasoning effort and clear_thinking are set on the server, so clients do not need custom request JSON. Keep speculative/MTP decoding off for this derivative. Increase context or concurrency only after a successful first run and checking actual free GPU memory.
Open a second PowerShell or Terminal window on your own computer. Copy the same full-SSH command from RunPod’s Connect panel and add “-N -L 127.0.0.1:8000:127.0.0.1:8000” after ssh, retaining its exact host, port, and identity-file options. Keep this tunnel window open. Your computer can now use http://127.0.0.1:8000/v1. No public HTTP port is required. If port 8000 is already used on your computer, forward local port 8001 instead and use that port in every client.
Check chat, tools, and streaming from your computer
Install Python 3.12 or newer on your own computer if it is missing. Download api-smoke-test.py to a new local folder, open PowerShell or Terminal in that folder, and run the command below. On macOS/Linux, use python3 if that is your Python command. The script checks model identity, chat with a follow-up, a real tool round-trip, and streaming, then saves api-check.json in this folder. Read its actual results; a server that only answers chat has not passed the tool check. The private tunnel setup uses a dummy client key unless you deliberately enable server authentication.
Put these files in Your local API test folder/1 required file
In a client supporting Chat Completions, choose an OpenAI-compatible connection and use the settings below. Keep the tunnel open. This is your private dealignai server; an OpenAI account or subscription is not needed for it. The server processes prompts, while your harness decides which local tools to expose and executes those tools on your computer. A successful API response alone is not an autonomous agent. For a visible coding task, continue with the optional Cline section below.
Setting
Value
API format
OpenAI-compatible Chat Completions
Base URL
http://127.0.0.1:8000/v1
API key
dummy for the private unauthenticated server; your chosen key if enabled
When finished, stop generation in your client. In the Pod Terminal, leave the log view with Ctrl+C and run the stop command below, then close the tunnel. This stops the model process, not RunPod billing: stop the Pod in RunPod’s console and confirm its state. To resume, start the Pod, reconnect, and repeat the background launch/PID/log commands from step 5 without downloading the setup again. Reopen the tunnel using current Connect details. Local coding files stay on your computer. Preserve needed server files before terminating a Pod or deleting storage.
This is the fastest setup to try before renting GPUs. MagmaRouter supplies its own GLM-5.3 derivative, not the dealignai weights above. Provider claims are source-checked; no availability, refusal-rate, or service-level guarantee is asserted.
01
Get a capped API key
Sign in to MagmaRouter, create an API key, and give it a small lifetime spending cap. Check current pricing and whether glm-5.3-xploded is available before adding more credit. No RunPod account, GPU rental, or local model download is needed. Prompts go to this hosted provider.
On Windows, macOS, or Linux, install VS Code from its official download page. Open Extensions, search for Cline, and install the linked extension. Open an empty local project folder. Install Python if you want to run the supplied standard-library coding task. These are local client tools; the model stays at the provider.
Open Cline Settings and choose OpenAI Compatible. Enter your own key and the values below. Start with 65,536 context and 8,192 output tokens for the small task, even though the provider advertises a larger window. Disable image input for this text-only model. These hosted settings replace the local tunnel settings; do not use the dealignai served alias here.
Select Act mode, keep tool-approval prompts enabled, and enter this task in Cline: “In this empty project, create calculator.py with an add(a, b) function and test_calculator.py using Python unittest. Test positive numbers, negative numbers, and zero. Run the tests, fix any failure, and tell me which files changed. Use only the Python standard library.” Approve the expected file edits and test command. Confirm that tools actually run; a prose answer claiming that files were created is insufficient.
In VS Code, open calculator.py and test_calculator.py and inspect their contents. Confirm the displayed unittest run passes, or rerun “python -m unittest -v” in the project terminal (python3 on macOS/Linux when needed). Stop the task and check usage against the key’s spending cap. The files remain in your local project folder. Reopen that folder and Cline to continue later; no GPU server needs restarting.
04
MORE WAYS TO USE THIS SETUP
Optional: use your private RunPod API in Cline
Use this after the exact self-hosted server passes the supplied API checks. Cline is the coding client; this section does not present a separate general-purpose agent workflow.
01
Install the local coding client
Install VS Code for Windows, macOS, or Linux and add the linked Cline extension. Open a new empty project folder on your computer. Install Python for the standard-library task. Keep the private SSH tunnel and remote model server running.
In Cline Settings choose OpenAI Compatible and enter the settings below. Disable image input. Leave extra reasoning options unset so the server’s low-effort default applies. Cline runs file and command tools in your local project; only inference runs on the Pod.
In Act mode, keep tool-approval prompts enabled and enter this task: “In this empty project, create calculator.py with an add(a, b) function and test_calculator.py using Python unittest. Test positive numbers, negative numbers, and zero. Run the tests, fix any failure, and tell me which files changed. Use only the Python standard library.” Inspect calculator.py and test_calculator.py in VS Code and confirm the unittest result passes. Review the code before using it. Your output lives in this local project folder; stopping the remote Pod does not delete these files.
05
MORE WAYS TO USE THIS SETUP
Hosted alternative: Abliteration.ai API
Abliteration.ai identifies abliterated-model-large-v2 as its GLM-5.3 derivative hosted in FP8. It is a different modification from dealignai and Xploded. Provider-documented setup checked October 23, 2026; live generation not tested by YouRunAI.
01
Get an API key and check availability
Open the Abliteration console, create a project API key, and check credit, spending limits, and access to abliterated-model-large-v2 before a longer run. Start with a small prepaid balance. No RunPod account or GPU rental is needed. The provider describes compute as limited and first come, first served, so a key does not establish reserved capacity. Prompts go to this hosted provider; platform protections and published project policies may apply.
Choose an OpenAI-compatible Chat Completions connection in your Windows, macOS, or Linux client and enter the values below. Use the explicit Large V2 ID: the generic abliterated-model and deprecated abliterated-model-large aliases select different models. Keep image input off. The provider advertises 1,000,000 total context tokens; the smaller first-task settings below are our starting choice. Your harness exposes and executes local tools; the API supplies model responses.
Setting
Value
API format
OpenAI-compatible Chat Completions
Base URL
https://api.abliteration.ai/v1
API key
Your Abliteration project key, beginning ak_
Model ID
abliterated-model-large-v2
Context / max output
65536 / 8192
Reasoning effort
low; top-level reasoning_effort in Chat Completions
For your own harness, POST this JSON body to https://api.abliteration.ai/v1/chat/completions with Content-Type: application/json and Authorization: Bearer YOUR_ABLITERATION_KEY. Use your own key locally. The low effort field is top-level; do not copy the RunPod-specific chat_template_kwargs or clear_thinking settings. Large V2 otherwise defaults to max reasoning and cannot turn reasoning off. Confirm streamed answer text arrives; the reasoning trace alone is not the final answer. The downloadable RunPod API checker is specific to that self-hosted route.
{
"model": "abliterated-model-large-v2",
"messages": [
{"role": "user", "content": "Give me a three-step plan for a small Python calculator with unit tests."}
],
"reasoning_effort": "low",
"max_tokens": 8192,
"stream": true
}
04
Optional: enter the same connection in Cline
Install VS Code and the linked Cline extension on your computer, then open an empty local project folder. In Cline Settings choose OpenAI Compatible, enter the Abliteration base URL, your key, and abliterated-model-large-v2 from the table above. Set 65,536 context and 8,192 max output, with image input disabled. If your client exposes reasoning effort, choose low; an unset effort uses the provider’s max default. Keep tool-approval prompts enabled. Install Python to run the coding task below.
In your coding harness, or Cline Act mode, enter: “In this empty project, create calculator.py with an add(a, b) function and test_calculator.py using Python unittest. Test positive numbers, negative numbers, and zero. Run the tests, fix any failure, and tell me which files changed. Use only the Python standard library.” Approve the expected file changes and test command. Inspect calculator.py and test_calculator.py in the local project folder and confirm a passing unittest run. A written claim without files and test output is insufficient. Stop the task and check the console’s usage and remaining credit. Local files remain on your PC; resume by reopening the project and reconnecting the client. No rented Pod needs shutting down.
For the exact self-hosted route, open the saved api-check.json and confirm the intended model ID, usable chat with a follow-up, a structured automatic tool call, its locally executed result, and streaming through the private tunnel. For a hosted route, confirm your client uses glm-5.3-xploded at MagmaRouter or abliterated-model-large-v2 at Abliteration.ai, creates the two local Python files, and runs passing tests. Record the selected build’s identity. A client error, partial response, or a claim of tool use without actual output does not pass.
RunPod and MagmaRouter sources checked October 22, 2026; the added Abliteration.ai option checked October 23. These instructions do not establish a successful GPU deployment, hosted generation test, or service reliability benchmark. “Uncensored” is a publisher/provider claim about refusal behavior. The exact dealignai checkpoint, MagmaRouter Xploded, and Abliteration Large V2 are distinct builds. Stop RunPod compute explicitly after use and review continuing storage charges.
The server cannot load the checkpoint or runs out of memory
Confirm one host contains all eight H200 141 GB GPUs, that /workspace has enough persistent free space, and that the pinned vLLM/CUDA build matches the driver. Keep MTP off. Start with the prepared 65,536 context and one active sequence; reduce context before increasing concurrency. A single H100/H200 or consumer GPU is not this route.
The PC cannot reach http://127.0.0.1:8000/v1
Check that server startup completed on the Pod and that full SSH uses its current public IP and mapped port. Keep the -L tunnel window open. Use the full SSH-over-exposed-TCP command, not the basic proxy. If local port 8000 is occupied, change only the local forwarded port and update every client URL.
The model thinks for a long time and returns empty content
Use reasoning_effort low, the prepared clear_thinking setting, and an 8,192-token output cap. On this checkpoint, off, medium, max, or an unset effort can fall back to maximum reasoning. A finish_reason of length with reasoning but no content indicates token-budget exhaustion. Current vLLM exposes reasoning in message.reasoning.
Chat works but tool calls appear as plain text or fail in Cline
Run the automatic tool-call check before Cline. Keep the checkpoint’s glm47 tool parser, glm45 reasoning parser, and auto tool choice enabled. The returned tool_calls arguments must parse, and the next answer must use the tool result. In Cline use the exact model alias, matching context, and OpenAI Compatible provider. Reconfigure the provider if an old capability setting disabled tools.
The hosted API rejects the model or key
Use https://api.magmarouter.com/v1 with glm-5.3-xploded and that provider’s own key. Check current model availability, account credit, and the key cap. This is a different build from dealignai; the private server alias glm-5.3-uncensored does not identify a hosted model.
Abliteration rejects a request or returns reasoning without an answer
Pair https://api.abliteration.ai/v1 with abliterated-model-large-v2 and your own ak_ project key. Check credit, limits, availability, and any project policy block in the console. Send text input only. Set top-level reasoning_effort to low; the default is max and hidden reasoning still uses tokens. Inspect finish_reason before increasing max_tokens. Hosted reasoning is returned in message.reasoning_content. Omit tools and tool_choice when no tools are supplied; an empty tools array is rejected.
The Pod is stopped but charges remain
Check compute state in RunPod and inspect separate volume/network-volume charges. Closing SSH or pressing Ctrl+C only stops your session or process. Preserve needed files before terminating a Pod; its attached Pod volume is deleted on termination, while network volumes are managed separately.