Summary
Business Insider profiles GLM-5.2, a new open-source large language model from
Chinese company z.AI (formerly Zhipu AI), engineered for long-horizon coding
and agentic workflows. It ships with a 1-million-token context window — the
same tier as Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 — and, like
DeepSeek before it, is freely downloadable and modifiable (MIT-licensed weights on
Hugging Face). The hook is the reaction: prominent Western technologists (Vercel's
Guillermo Rauch, ex-Meta/Google/Microsoft VP Matt Velloso) publicly called it the
first open model good enough to be a daily driver for coding. The piece frames it
as the next "DeepSeek moment" in the US–China AI race: Washington leans on chip
controls to preserve a lead while Chinese labs ship cheaper, increasingly capable
open models that erode the value advantage of closed American frontier systems.
Key points
- What it is: GLM-5.2 — an open-weights LLM from z.AI (formerly Zhipu AI),
Beijing — purpose-built for extended coding tasks and agentic workflows. - Context window: 1,000,000 tokens, putting it in the same league as Claude
Opus 4.8 and GPT-5.5 for long-context work. - Genuinely open: released under an MIT license with weights on Hugging
Face — download, fine-tune, and self-host with no usage restrictions. - Architecture (per secondary sources): a Mixture-of-Experts model, ~744–753B
total parameters with only ~40B active per inference — big capacity, cheaper to run. - Beats GPT-5.5 on coding, far cheaper: reported to top GPT-5.5 on multiple
long-horizon coding benchmarks (e.g. SWE-bench Pro 62.1 vs 58.6) at roughly
1/6th the cost. - The reactions are the story: Guillermo Rauch (Vercel CEO) — "Genuinely
impressed, almost shocked, at how good GLM-5.2 ... is at coding. This changes
things." Matt Velloso (ex-Meta/Google DeepMind/Microsoft VP) — "First open model
that passes the bar as a daily driver." - Geopolitical frame: positioned as another DeepSeek-style signal — US chip
restrictions vs. China's cheap, capable, open-source strategy that undercuts the
closed-model business case. - Why it matters for us: an MIT-licensed, self-hostable, frontier-ish coding
model with 1M context is directly relevant to agentic dev tooling — worth tracking
as a cost/sovereignty alternative to closed APIs.
Fact check
| Claim | Verdict | Notes | Source |
|---|---|---|---|
| GLM-5.2 has a 1M-token context window, same tier as Claude Opus 4.8 & GPT-5.5 | ✅ Confirmed | 1M context corroborated across multiple trackers; Claude Opus 4.8 does offer a 1M-token context tier | VentureBeat |
| Made by z.AI, a Chinese company (formerly Zhipu AI), Beijing | ✅ Confirmed | z.AI is the rebranded Zhipu AI | Requesty |
| Open-source — anyone can download and modify it | ✅ Confirmed | Core weights released under an MIT license on Hugging Face | VentureBeat |
| Beats GPT-5.5 on multiple long-horizon coding benchmarks | ✅ Confirmed | e.g. SWE-bench Pro 62.1 vs GPT-5.5 58.6 | VentureBeat |
| Costs roughly 1/6th of GPT-5.5 | ⚠️ Partially Confirmed | VentureBeat headline figure; effective ratio depends on input/output mix & caching (~$1.40/M in, $4.40/M out) | Requesty |
| ~744–753B-parameter MoE with ~40B active | ✅ Confirmed | Reported as a 753B-total / 40B-active Mixture-of-Experts model | CloudPrice |
| Guillermo Rauch (Vercel CEO) praised it as a coding game-changer | ✅ Confirmed | Direct quote reproduced in the Business Insider article | Business Insider (Varanasi, 2026-06-21) |
| Echoes the DeepSeek R1 moment (released Jan 2025) | ✅ Confirmed | DeepSeek R1's Jan 2025 release is the stated precedent | Business Insider (Varanasi, 2026-06-21) |
Related resources
| Type | Name | URL | Notes |
|---|---|---|---|
| model | GLM-5.2 (z.AI) on Hugging Face | https://huggingface.co/zai-org | MIT-licensed open weights |
| guide | Unsloth — Run GLM-5.2 locally (GGUF quants) | https://unsloth.ai/docs/models/glm-5.2 | 2-bit dynamic GGUF (~240 GB) + setup |
| guide | Spheron — Deploy GLM-5.2 on GPU cloud | https://www.spheron.network/blog/deploy-glm-5-2-gpu-cloud/ | VRAM tiers + rental cost |
| pricing | OpenRouter — GLM-5.2 providers & pricing | https://openrouter.ai/z-ai/glm-5.2 | Live API price comparison across hosts |
| vendor | z.AI (formerly Zhipu AI) | https://z.ai | Beijing-based developer of the GLM family |
| analysis | VentureBeat — GLM-5.2 beats GPT-5.5 for 1/6th the cost | https://venturebeat.com/technology/z-ais-open-weights-glm-5-2-beats-gpt-5-5-on-multiple-long-horizon-coding-benchmarks-for-1-6th-the-cost | Benchmark + pricing write-up |
| context | DeepSeek R1 (Jan 2025) | https://www.deepseek.com | The earlier "open Chinese model turns heads" moment the article echoes |
Full content
(click to expand)
GLM-5.2 is an open-source large language model from the Chinese company z.AI
(formerly Zhipu AI), built for long coding tasks and agentic workflows. It runs on
a 1-million-token context window — the same tier as Anthropic's Claude Opus 4.8 and
OpenAI's GPT-5.5 — and, like DeepSeek, is open-source: anyone can download it, run
it on their own systems, and modify it. The model launched to a wave of attention
from investors, founders, and industry figures: Vercel CEO Guillermo Rauch said he
was "almost shocked" at how good it is at coding, and former Meta/Google
DeepMind/Microsoft VP Matt Velloso called it "the first open model that passes the
bar as a daily driver." Independent write-ups report it beating GPT-5.5 on several
long-horizon coding benchmarks at roughly one-sixth of the cost, under an MIT
license with weights on Hugging Face. The article situates GLM-5.2 in the US–China
AI contest: Washington restricts chips to hold its edge, while Chinese labs push
cheaper, increasingly capable open models that pressure the closed-model business
case of OpenAI and Anthropic.