GLM-5.2

Z.ai's (formerly Zhipu) flagship open-weights model, released 2026-06-17: a ~753B-parameter Mixture-of-Experts (~40B active) tuned for long-horizon coding/agent tasks, with a usable 1M-token context. The strongest open-source coding model at launch, within ~1% of Claude Opus 4.8 on several long-horizon benchmarks at roughly 1/6th the cost. Source: VentureBeat, 2026-06-17; Z.AI docs, docs.z.ai/guides/llm/glm-5.2, 2026-06-17

What it is

GLM-5.2 is built "for the era of long-horizon tasks" — the thesis is not just a 1M context but a usable one: stable performance and engineering-standard adherence across multi-hour, cross-file, project-scale work. A single task is meant to carry an entire codebase's module boundaries, API contracts, and historical decisions from requirements through to a deployable product. Source: Z.AI docs, 2026-06-17

Release shape was unusual: a 2026-06-13 soft launch to GLM Coding Plan subscribers, then full MIT-licensed open weights + the first official benchmark table on 2026-06-17 (Hugging Face + ModelScope). Source: TechTimes, 2026-06-17

Kevin bookmarked the launch the day it dropped. Source: X/@Zai_org, 2026-06-17

Architecture

• Mixture-of-Experts, ~753B total parameters with ~40B active per token (a learned gate routes each token to the top-2 of hundreds of experts), so it stays deployable despite the size. vLLM's recipe lists it as ~743B / 39B active — treat the headline number as ≈753B total / ~40B active. Source: TechTimes, 2026-06-17; vLLM Recipes, recipes.vllm.ai/zai-org/GLM-5.2
• IndexShare attention reportedly cuts per-token compute by ~2.9× at the full 1M context, which is what makes the long window practical rather than nominal. Source: VentureBeat, 2026-06-17
• Multi-Token Prediction (MTP) extended from 3 → 5 draft tokens vs GLM-5/5.1, lifting end-to-end throughput on reasoning/coding/agentic workloads via speculative decoding. Throughput is +3–192% over GLM-5.1 across 32K–1M context, with the largest gains at the longest contexts. Source: vLLM Recipes; TechTimes, 2026-06-17
• Ships as BF16 and native-FP8 checkpoints. FP8 fits a single 8×H200 / 8×H20 node; the full 1M context is reachable on 8×B200 with an FP8 KV cache. Served via vLLM 0.23. Source: vLLM Recipes, recipes.vllm.ai/zai-org/GLM-5.2

Benchmarks

State-of-the-art among open-source coding models at launch; closing on the closed frontier. Source: Z.AI docs + VentureBeat, 2026-06-17

The reviewed launch image is a vendor benchmark chart for "Agentic Coding Performance by Effort Level," averaged over Terminal-Bench 2.1, DeepSWE, and SWE-Atlas QnA, evaluated on Claude Code 2.1.167. It shows GLM-5.2 improving sharply from non-thinking to high effort, then gaining only modest score at max effort while output tokens rise toward ~90K/task. Treat the image as useful source evidence for the effort/cost tradeoff, not as an independent benchmark. Source: X/@Zai_org, 2026-06-17

The Vercel follow-up adds a current agent-eval anchor rather than another vendor launch chart. On the Next.js AI Agent Evaluations page, the 2026-06-09 run lists GLM 5.2 via OpenCode at 281.81s, 88% success, and 96% success with bundled AGENTS.md documentation. The durable signal is twofold: GLM-5.2 is competitive among open coding agents on a framework-specific suite, and agent-facing project documentation changes the pass rate enough to be treated as part of the runtime, not mere repo garnish. Source: Next.js evals, checked 2026-07-04; Source: X/@vercel_dev, 2026-06-16

The Proximal FrontierSWE bookmark adds a stronger independent benchmark anchor than the original launch post. The live FrontierSWE page ranked GLM-5.2 #3 on mean@5 as of 2026-06-30: avg rank 4.32, dominance 74%, behind Claude Fable 5 and Claude Opus 4.8, ahead of GPT-5.5. The reviewed video also shows the best@5 view, where GLM-5.2 moves to #2 behind Claude Fable 5, and the thread notes its best research-task result on PCQM4Mv2 Molecular Gap Prediction. Keep both modes because they answer different questions: mean@5 is stability across trials; best@5 is peak capability under retries. Source: FrontierSWE, 2026-06-30; Source: X/@ProximalHQ, 2026-06-16

Theo's reviewed cost plot is a useful counterweight to "open weights means cheap." In the DeepSWE cost view, glm-5.2 [max] sits below gpt-5.5 [medium] and claude-opus-4.8 [high] on score while also landing in a less efficient cost band than those medium/high closed-model routes. His self-thread adds the operational detail: cheaper tokens can be erased by higher output-token volume and slower wall-clock time. Use GLM-5.2 where open weights, data control, or long context matter; do not route it as the default cheap frontier substitute. Source: X/@theo, 2026-06-21

BridgeMind's reviewed screenshot adds a dated BridgeBench anchor: GLM-5.2 was shown at #1 on the BS benchmark with 100.0 and #1 on Reasoning with 42.8, ahead of Claude Fable 5 on that board. The live BridgeBench pages still show GLM-5.2 at 100.0 on BS and 42.8 on the 30-task reasoning suite. Treat this as benchmark-specific evidence for open-weights competitiveness, not a universal claim that GLM-5.2 beats Fable 5 everywhere. Current Anthropic docs also changed the availability backdrop: Fable 5 access was restored starting 2026-07-01 after the June suspension. Source: X/@bridgemindai, 2026-06-14; Source: BridgeBench BS, checked 2026-07-03; Source: BridgeBench Reasoning, checked 2026-07-03; Source: Anthropic, 2026-07-01

Capabilities & API

• Two reasoning-effort levels: GLM-5.2 (max) pushes limits; GLM-5.2 (high) is the strong default. Thinking mode is toggleable (thinking: {type: enabled|disabled}, reasoning_effort: max). Source: X/@Zai_org, 2026-06-17; Z.AI docs
• Tool calling, MCP integration, JSON structured output, prompt caching, real-time streaming (with separate reasoning_content deltas). Source: Z.AI docs, 2026-06-17
• OpenAI-compatible API: base https://api.z.ai/api/paas/v4/, model="glm-5.2". Official zai-sdk (Python, Java) and the OpenAI SDK both work. Source: Z.AI docs, 2026-06-17
• Pricing: ~$1.40 /M input, $4.40 /M output (mid-priced globally, ~1/6th of GPT-5.5); GLM Coding Plan subscriptions from ~$12.60/mo. Source: VentureBeat, 2026-06-17

Where to run it

Open weights on Hugging Face + ModelScope; hosted on the Z.ai API, Ollama cloud (US, NVIDIA Blackwell), Vercel AI Gateway (model: 'zai/glm-5.2'), and 20+ third-party coding environments (Kilo Code confirmed day-one). The Vercel changelog confirms the 1M-token upgrade from GLM-5.1's 200K window, the AI SDK model id, and Gateway features that matter in production routing: usage/cost tracking, retries, failover, ZDR support, API-key budgets, BYOK, and no platform markup on inference. The live AI Gateway model page lists multiple providers behind the same model slug, including Z.ai, Baseten, Fireworks, Wafer, DeepInfra, and Parasail; provider latency/throughput/pricing should be treated as live routing data, not a static model property. The Ollama model page repeats the important routing claim: "solid 1M context" for long-horizon coding and agentic tasks, high/max effort modes, and MIT open weights. A later Ollama support thread clarified that GLM-5.2 cloud capacity was doubled on US-based NVIDIA B300 Blackwell GPUs, while the privacy-policy replies say local Ollama runs are unseen by Ollama and cloud-hosted prompts/responses are processed transiently, not used for training. Source: Vercel changelog, 2026-06-16; Source: Vercel AI Gateway model page, checked 2026-07-04; Source: X/@vercel_dev, 2026-06-16; Source: X/@ollama, 2026-06-16; Source: X/@ollama, 2026-06-18; Source: Ollama privacy policy, 2026-06-30; Source: Ollama model page, 2026-06-30

Current Hugging Face snapshot as of 2026-07-04: model zai-org/GLM-5.2, public/non-gated, SHA b4734de4facf877f85769a911abafc5283eab3d9, last modified 2026-07-02 08:08 UTC, 3,345 likes, and 191,462 downloads. Source: Hugging Face API, 2026-07-04

Because it is on AI Gateway, it is a drop-in model choice for a Vercel Eve agent (model: "zai/glm-5.2") and for anything routing through Vercel AI SDK.

ZCode — Z.ai's coding agent

zcode.z.ai hosts ZCode 3.0, Z.ai's desktop coding-agent IDE deeply adapted to GLM-5.2 (its top effort tier shows as "GLM-5.2 Max"). It is a multi-agent harness in the Cursor / Claude Code / Loop category: cross-repo codebase understanding, automated inline code review before PRs, a command palette + integrated terminal + workspace, and integration with existing editors, task systems, and release pipelines (it ships demo agents like a release-bot). Downloads for macOS (Apple Silicon), Windows, and Linux (Beta). Source: zcode.z.ai, 2026-06-17

Relevance to Kevin

• Open weights = self-hostable on Dedalus Labs microVMs / Machines, sidestepping the API "China data residency" caveat flagged for the hosted endpoint. Source: TechTimes, 2026-06-17
• A long-horizon coding model directly comparable to the Opus/GPT options in Kevin's harness work; the reasoning-effort + thinking-mode knobs are a concrete instance of language/compute as an orchestration variable (Language Orchestration / CoTCodec).
• Reinforces the 2026 pattern (cf. X For You Algorithm (2026 Open Source)): frontier coding capability shipping as open weights, usable behind a gateway alongside Western models.

---

Timeline

• 2026-07-04 | Source-reviewed @vercel_dev's AI Gateway bookmark. Added the official Vercel changelog/model-page route, Next.js eval row (GLM 5.2 via OpenCode at 88% / 96% with AGENTS.md), and refreshed Hugging Face stats while the SHA remained b4734de. Source: X/@vercel_dev, 2026-06-16; Source: Next.js evals, checked 2026-07-04
• 2026-07-03 | Deep-reviewed @bridgemindai's BridgeBench screenshot. Added the local image artifact plus live BridgeBench BS/Reasoning anchors: GLM-5.2 at 100.0 on BS and 42.8 on Reasoning. Also corrected the Fable availability context: Anthropic redeployed Fable 5 starting 2026-07-01, so the "banned" framing is historical. Source: X/@bridgemindai, 2026-06-14; Source: BridgeBench, checked 2026-07-03; Source: Anthropic, 2026-07-01
• 2026-07-02 | Deep-reviewed Theo's GLM-5.2 cost/throughput bookmark and repaired its local image manifest. Added the DeepSWE cost plot caveat: GLM-5.2 max is impressive but not automatically cheaper once output-token volume and wait time are included. Refreshed Hugging Face to SHA b4734de, 3,194 likes, and 159,967 downloads. Source: X/@theo, 2026-06-21; Source: Hugging Face API, 2026-07-02
• 2026-06-30 | Deep-reviewed the Design Arena bookmark and local chart artifact. Added the source card and promoted Design Arena as the benchmark owner page so the 1360 Elo claim stays a dated snapshot rather than a permanent routing fact. Source: X/@Designarena, 2026-06-16; Source: Design Arena, 2026-06-30
• 2026-06-30 | Deep-reviewed the second Ollama GLM-5.2 cloud bookmark. Added the capacity/privacy nuance: doubled US-based B300 Blackwell capacity for GLM-5.2 cloud, with Ollama's privacy policy as the durable source for local-vs-cloud prompt handling. Source: X/@ollama, 2026-06-18; Source: Ollama privacy policy, 2026-06-30
• 2026-06-30 | Deep-reviewed the Proximal FrontierSWE bookmark and repaired its local thumbnail artifact. Added the independent FrontierSWE ranking nuance: GLM-5.2 is #3 on mean@5, #2 on best@5 in the reviewed video, and strongest on the PCQM4Mv2 research task. Refreshed Hugging Face likes to 3,002 while the SHA remained f226310. Source: X/@ProximalHQ, 2026-06-16; Source: FrontierSWE, 2026-06-30
• 2026-06-30 | Deep-reviewed the Ollama cloud bookmark and repaired the local media artifact. Added the Ollama source card and root tweet; the reviewed image is a Z.ai vendor benchmark grid across SWE-bench Pro, Terminal-Bench 2.1, NL2Repo, DeepSWE, ProgramBench, MCP-Atlas, Tool-Decathlon, and Humanity's Last Exam, all at maximum thinking effort. Treat it as vendor evidence for breadth, not independent proof. Source: X/@ollama, 2026-06-16
• 2026-06-30 | Deep-reviewed launch image and current HF source. Added the benchmark-image caveat, local image artifact, and HF snapshot (f226310, MIT, 2,973 likes, 142,547 downloads). Source: X bookmark artifact audit, 2026-06-30
• 2026-06-17 | Captured from Kevin's X bookmarks ("pull my x bookmarks … super detailed entries about glm 5.2"). Open weights (MIT) + first benchmark table released; #1 on Design Arena (Elo 1360); available on Z.ai API, Ollama cloud, Vercel AI Gateway (zai/glm-5.2), HF/ModelScope. Noted ZCode 3.0 (zcode.z.ai) and docs.z.ai/guides/llm/glm-5.2. Param count ≈753B total / ~40B active (vLLM lists ~743B/39B). Source: User, 2026-06-17; Z.AI docs; VentureBeat; TechTimes; vLLM Recipes; X/@Zai_org, @vercel_dev, @ollama, @Designarena, 2026-06-17
• 2026-06-13 | Soft-launched to GLM Coding Plan subscribers ahead of the open-weights release. Source: TechTimes, 2026-06-17
• 2026-06-14 | Positioned as the open-weights answer to US export restrictions: @bridgemindai noted that two days after the US "banned" Claude Fable 5, GLM-5.2 topped @bridgebench (BS 100.0, Reasoning 42.8 — beating Fable 5) at ~1/10th the cost. Reinforces the open-weights-vs-frontier-lab dynamic. Source: X/@bridgemindai, 2026-06-14