Sponsored by
This week you can download a coding model that lands within 4 points of Opus 4.8 on Terminal-Bench, point it at your own GPUs, and never send a line of your code off your network.
It is free, MIT licensed, and it has been on Hugging Face since the 16th. I spent two days convinced this was the release that made the API obsolete, then I priced the hardware.
GLM 5.2 is the real thing, a frontier-adjacent model you can own outright instead of rent, with no per-token meter and no data leaving the building.
The catch is what it takes to run.
Once you have seen the GPU bill, most of you are going to want to keep renting.
If you ran an open coding model yourself, what would push you to do it?
Before We Begin
DevTools Of The Week
Locofy: Converts Figma and Penpot designs into front-end code across React, React Native, Vue, Angular, Next.js, Flutter and plain HTML-CSS. It is SOC2 and ISO certified, and can be deployed on-prem or in the cloud.
refuse: An open-source shim that wraps npm, pip, cargo, gem, go and 13 other package managers, checking every install against 362,000+ advisories across 21 ecosystems and blocking vulnerable versions whether you ran the install, your agent did, or CI did.
Otty: Otty is a native, GPU-accelerated macOS terminal built for code agents, with first-class support for running Claude Code, Codex and OpenCode side by side, a live web preview that updates as the agent edits, and the option to pipe any output straight into the chat.
What You Are Downloading
GLM 5.2 is a 744B parameter Mixture-of-Experts model, 40B active per token.
The full BF16 checkpoint on Hugging Face runs 1.51 TB and the FP8 build (zai-org/GLM-5.2-FP8) is roughly 750 GB.
The architecture is where the long-context story holds up.
On top of DeepSeek Sparse Attention, Z.ai added a trick called IndexShare that reuses one indexer across every 4 sparse-attention layers and cuts per-token FLOPs by 2.9x at 1M context.
The MTP layer for speculative decoding improved as well, up to 20% higher acceptance length. Long context ends up cheaper to serve here than on a dense model, and that part holds up under a real repo load.
Running everything else is where the money goes.
The Hardware Reality
The trap is the active-parameter number.
GLM 5.2 has 744B parameters but fires only 40B per token, so people size their GPUs for 40B and wonder why nothing loads. The whole model has to sit in memory at once, whether a given token touches an expert or not.
Size for the full 744B, always.
From there it is a question of precision, which is just how many bits you spend storing each number. Drop bits and the model shrinks while accuracy slips a little.
Three ways to play it:
FP8 (~750 GB): fits on 8 H200s, leaves ~380 GB for the running conversation. The standard pick.
BF16 (~1.5 TB): double the size, needs ~16 GPUs. Overkill, skip it.
4-bit / INT4 (~200 GB): runs on 4 H200s or 8 cheaper H100s, costs you 1 to 3 percent on coding benchmarks. The budget way in for a personal agent.
Skip the community NVFP4 conversions floating around. They are untested and not production-ready.
Serving it: vLLM or SGLang
Both expose an OpenAI-compatible endpoint and both work fine and they pull apart during the workload.
Start with vLLM, it is the default for a reason. FP8 wants DeepGEMM installed first.
This serves at http://localhost:8000/v1. The glm47 tool parser and glm45 reasoning parser are not optional here.
Leave them off and both tool calling and the thinking-mode split break.
SGLang earns its place once you are doing agentic work, and the reason is RadixAttention.
Every agent step resends the same big codebase plus one new instruction, and RadixAttention keeps that codebase KV cached so you only pay for the new portion.
Across a long multi-turn loop that adds up to real throughput.
--enable-moe-ep is the expert-parallel flag. Without it, tensor parallelism copies every expert to every GPU and wastes the memory you just paid for.
A warning on context, because this one bites people: do not run the full 1M by default.
For agentic coding your practical window is 32K to 131K.
Save the full million for whole-codebase ingestion, and when you do reach for it, FP8 KV cache stops being optional or you will blow your VRAM the moment you batch.
Wiring It Into A Coding Agent
Skip Claude Code for this one.
It speaks the Anthropic Messages API while your vLLM or SGLang server speaks OpenAI, so bridging them means parking a translation proxy like claude-code-router or LiteLLM in front.
That is a moving part you do not need.
Reach instead for an agent that talks OpenAI natively.
OpenCode is the cleanest fit, CLI based, multi-provider, and it points at any OpenAI-compatible endpoint in a few lines of config.
If you would rather stay in VS Code, Cline and Kilo Code get you there through their custom-provider settings.
The OpenCode setup, start to finish:
Step 1: Confirm the server is answering.
Before touching any config, make sure the model responds on the OpenAI route. Use the --served-model-name you set when you launched vLLM.
You should see glm-5.2-fp8 in the list. If you get nothing, fix the server first.
No config will save a dead endpoint.
Step 2: Create the OpenCode config.
Open ~/.config/opencode/opencode.json for a global setup, or drop an opencode.json in your project root to scope it to one repo. Project config wins over global.
Two things that break people here.
The npm value must be @ai-sdk/openai-compatible, that is what tells OpenCode to use the OpenAI wire format, and the model key (glm-5.2-fp8) must match the served model name exactly.
A mismatch returns a NotFoundError that looks like a network problem but is not.
Step 3: Add a credential.
vLLM does not validate local keys, but OpenCode still wants one attached to the provider. Run:
Choose Other, enter glm-local as the provider ID, and paste any non-empty string like sk-local. It gets stored in ~/.local/share/opencode/auth.json.
The provider ID has to match the one in your config or the credential attaches to nothing.
Step 4: restart and select the model.
OpenCode does not hot-reload provider config.
Quit it fully, relaunch, then run /models and pick glm-local/glm-5.2-fp8.
Step 5: sanity-test before real work.
Send one small prompt and confirm a tool call fires.
Local models are pickier about tool-call formatting than hosted ones, which is why the glm47 and glm45 parsers on the server side matter. If tool calls silently fail, that is the first place to look.
That is the whole loop.
My Take
I went in thinking the headline was cost and I was wrong. For almost everyone reading this, the cost math is a reason not to do this at all.
An 8x H200 node bills the same whether it is slammed or sitting idle. You only claw that back if your token volume is high and your batching is tight, basically near-constant load.
A solo builder with spiky usage will pay less on GLM's $18 Coding Plan and spend 0 weekends babysitting vLLM.
I ran the numbers expecting it to be closer, but it wasn't.
There is one group this is built for: teams with proprietary code they cannot hand to a hosted Chinese API.
That is the real buyer.
The China data-law exposure is a live compliance problem for a lot of teams, and an MIT-licensed model running air-gapped on your own boxes makes it vanish.
If that is you, GLM 5.2 is the strongest thing you can own outright right now.
If it is not, you are renting an 8-GPU node to feel like an infra team.
The long context surprised me more than the benchmarks did.
I have watched a dozen 1M-context claims fall over the moment you load a real repo into them, and this one holds.
IndexShare and the FP8 KV cache are doing real work, and a year ago this was the exact thing that broke first.
Worth watching even if you never rack a single GPU.
My honest call is simple: run the break-even before you open a GPU rental page, and for most of you it is going to point straight back at the API.
Until next time,
Vaibhav 🤝🏻
If you read till here, you might find this interesting
#Partner 1
Six Tools, Zero Efficiency. There's a Better Way.
Most e-commerce sellers are running their store across 6 to 8 separate tools — switching tabs, reconciling data, and paying hundreds a month for the privilege. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.
It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.
Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.
Free to start. No credit card required.
#Partner 2
Stop switching apps. Your browser can do it all.
Every tab you open, every copy-paste into ChatGPT, every lost train of thought — that's your browser failing you. Norton Neo fixes it. Built-in AI works directly inside your session. Hover to preview. Search everything from one bar. VPN and ad blocking included, free.
