SHELFCRITTER AI Digest

MODEL

GLM 5.2

modeltopic-noteopen-sourcechinaz-ai

Overview

GLM 5.2 is Z.ai‘s June 2026 frontier open-weights release — a 744B-parameter mixture-of-experts model with a 1M-token context window, dual thinking-effort modes (“fast” and “deep”), and an MIT open-weights license scheduled for release the week after launch. The model is live across Z.ai’s GLM Coding Plan tiers and an API and chatbot opened the same day as the announcement. GLM 5.2 succeeds GLM-5.1 in the Chinese open-weights frontier cohort that includes DeepSeek V4 and the Qwen 3.x family.

Timeline

  • 2026-06-14-AI-Digest — Z.ai releases GLM 5.2 on 2026-06-13 with co-founder Jie Tang announcing the drop on X. Headline specs: 744B-parameter mixture-of-experts, 1M-token context window, dual thinking-effort modes (a “fast” pass and a “deep” pass), and an MIT-licensed open-weights release scheduled for next week alongside an API and chatbot opening today. Marketing emphasises coding and long-horizon agent use. Z.ai published no benchmark numbers at launch — not Aider, not SWE-Bench, not MMLU, not even an internal eval card; AI Weekly flagged the omission explicitly. Treat the headline as a release event, not a leaderboard event.
  • 2026-06-18-AI-Digest — GLM 5.2 now sits at the top of Artificial Analysis’s open-weights ranking and #4 overall on the Intelligence Index (score 51) — the leaderboard print arrives this week. The structural read: Chinese labs have held the top open-weights slot continuously through Q2 2026 (rotation Kimi K2.6 → DeepSeek V4 Pro → MiMo-V2.5 → GLM-5.1 → GLM 5.2 across ~6 weeks). The disciplined caveat is the agentic-coding axis: today’s Aider polyglot top-5 is sweeps-of-GPT-5 plus o3-pro and Gemini 2.5 Pro — no open-weights entry on the polyglot bar — but on frontend coding specifically Simon Willison flags GLM 5.2 as the new leader per a Latent Space note. The day’s juxtaposition: open-weights leadership is real and accelerating on general intelligence and on certain coding axes (frontend), while still trailing on the agentic-polyglot bar.
  • 2026-06-20-AI-Digest — GLM 5.2 referenced in the Aider-polyglot ten-days-frozen callout as the continuing open-weights leader on Artificial Analysis even as the polyglot board stays wall-to-wall closed reasoning. No fresh release or benchmark today; the durability across a ten-day window is itself the signal — two coding axes (frontend vs polyglot), two leaderboards, two leaders.
  • 2026-06-23-AI-Digest — Two same-day reference points sharpen the GLM 5.2 picture. (1) Unsloth guide for running GLM 5.2 locally hits the HN front page (271 pts / 129 cmts) with quantization and inference recipes — the front-page traction matches the broader sentiment moment on open-weights deployability and lands inside today’s three-open-weights-wins HN cluster (alongside Moebius 0.2B and VibeThinker 3B). (2) External coverage flags GLM 5.2 claiming wins against GPT-5 on SWE-bench Pro and Terminal-Bench 2.1, suggesting the corpus’s running “frozen-polyglot frame” may be eval-specific rather than capability-wide. The corpus will continue carrying the polyglot-vs-Pro axes separately until one of them speaks to the other; the Aider polyglot has now been frozen for thirteen days while sentiment on open-weights deployability is visibly moving.
  • 2026-06-29-AI-Digest — GLM 5.2 beats Claude on Semgrep’s internal IDOR cyber sub-task — Semgrep’s blog post (612 pts · 298 cmts on HN, titled “We Have Mythos At Home: GLM 5.2 Beats Claude in Our Cyber Benchmarks”) reports GLM 5.2 outscoring Claude Code narrowly on the IDOR sub-task with 39% F1 vs Claude Code’s 32%, with no scaffolding. The corpus framing the digest carries with precision: narrow and one benchmark, not generalized parity — today’s Aider polyglot top-5 still contains zero open-weights entries at day nineteen of the freeze, so GLM 5.2 is reaching parity on a single Semgrep cyber sub-task, not on broad agentic coding. Another data point that open-weights Chinese frontier models are closing on closed US labs on narrow specialist evals; the secondary axis watching is whether GLM 5.2 surfaces outside the polyglot top-5 cut on the next print.
  • 2026-07-02-AI-Digest — Z.ai‘s ZCode coding-agent harness for GLM 5.2 launches publicly at zcode.z.ai and hits the HN front page (306 pts / 248 cmts, title + URL only, story_text_len=0). The public harness launch extends the earlier API/chatbot distribution surface with a coding-agent-specific product. Reinforces the Chinese-open-weights coding-agent cadence read alongside LongCat-2.0 and the prior 30-day cluster; the day-twenty-two Aider polyglot freeze remains the contrasting axis — GLM 5.2 has not yet surfaced on the polyglot top-5, so ZCode is a distribution/harness event on the open-weights side rather than a polyglot-leaderboard event.
  • 2026-07-04-AI-Digest — HN vendor-blog headline: “GLM 5.2 on AMD MI355X at 2626 tok/s/node at over 2× lower cost than Blackwell” (163 pts / 49 cmts, wafer.ai blog, story_text empty). Vendor blog claims GLM 5.2 inference on AMD’s MI355X hits 2626 tok/s/node at >2× lower cost per token than NVIDIA Blackwell. Narrow read: concrete price/perf datapoint feeding the AMD-vs-NVIDIA inference debate. Structural read the digest carries: carry as vendor-blog claim (not independently benchmarked) rather than a settled number — the pattern of Chinese open-weights frontier models used as reference workloads in AMD-vs-NVIDIA vendor-marketed inference comparisons continues, and the vendor-marketing-not-third-party discipline is load-bearing. The next signal is whether an independent benchmark from a non-vendor party reproduces the 2626 tok/s/node result at a comparable cost multiple.
  • 2026-07-07-AI-Digest — Martin Alderson’s “GLM 5.2 and the coming AI margin collapse” post hits HN at 254 pts / 165 cmts arguing Zhipu’s GLM 5.2 open weights match frontier quality at a fraction of the price, compressing inference margins for closed labs. Practitioner-framing signal, not a benchmark event — but the read pairs cleanly with today’s Alibaba / Claude Code ban as the same “open-weight domestic-Chinese stack looks materially more attractive” pressure surface hitting Western closed-lab commercial positioning from two sides in the same news cycle.
  • 2026-07-08-AI-Digest — GLM 5.2 anchors Zhipu AI‘s ZCode coding-agent launch, positioned explicitly against Claude Code and OpenAI Codex — 1M-token context, five-day trial of 5M tokens/day (3M GLM 5.2 + 2M GLM-5-turbo), paid plans starting $18/month, API pricing at $1.40 / $4.40 per M in/out (~1/6th of GPT-5.5). The Decoder cites a Snowflake CEO write-up of a 103-task dbt-bench comparison in which GLM 5.2 and Claude Opus 4.7 land 66% vs 67% at Pass@3 — but with a wider first-attempt gap (47.6% vs 53.7%) and roughly 2× the token usage on the GLM 5.2 side. Narrow read: on one SQL-coding benchmark at three attempts near-parity, but the Pass@1 gap and 2× token cost tell a different story about single-shot reliability and inference economics. Structural read the digest carries: the pricing is the news, not the benchmark — pairs with today’s Tencent Hy3 Apache 2.0 open-weights release as two independent pressure points on the coding-agent cost stack in one week.

Key Developments

  1. 744B MoE with 1M Context under MIT License: The release shape matches the playbook that put DeepSeek V4 and Qwen 3.x at the open-weights frontier — a frontier-scale MoE with a long context window under a permissive open license, with API/chatbot live for direct use and weights to follow.

  2. Benchmarks Withheld at Launch: The absence of any published benchmark numbers — not even an internal eval card — is the load-bearing discipline test. Without numbers, “frontier-closing” framings are vendor narrative; the practitioner hinge is whether independent evals next week confirm coding parity with GPT-5 / Claude Opus 4.8 tiers or land closer to the GLM-5.1 cohort.

  3. Dual Thinking-Effort Modes: A “fast” and “deep” pass on the same model continues the multi-effort-tier pattern that frontier labs have been formalising through 2026 (Opus 4.7’s xhigh tier, Gemini 3 Deep Think, GPT-5’s reasoning effort settings).

See also: Z.ai, GLM-5.1, DeepSeek v4, Qwen, Claude Opus 4.8, GPT-5, MOC - Open Source Models.