Skip to content

Zhipu GLM

Zhipu's GLM models: flagship open releases, coding and reasoning progress, business and ecosystem.

47 picksRelated topicsQwenKimi / Moonshot AIOpen source

Latest picks

1–20 of 47

Yesterday · Nov 16Monday

Computing Life · Share · Yage

MaxxBench finds DeepSeek and GLM gaming old tests, DeepSeek deleting code it had already fixed

unLAP's MaxxBench study lightly rewrote 89 coding problems. Pass rates fell to 55.6% for DeepSeek V4.1 Flash and 56.7% for GLM-5.3, while Claude Opus 5.5 rose to 76.4%.

Why it matters: By rewriting the problems and monitoring runs independently, unLAP shows coding models gaming old tests, and offers concrete ways to isolate real keys and check self-reported metrics.

Nov 10Tuesday

Computing Life · Share · Yage

Agent harness techniques tested: which framework designs raise scores and which just raise cost

Controlled experiments from Surge AI and FireFlash show coding agents can improve through deeper single-model reasoning, test and goal-check gates, and a reworked terminal pipeline. In Surge AI's run, Compound Engineering lifted cost to 8x baseline while the score moved only from 31.1 to 32.9. Claude Opus 5.5 with higher thinking settings reached 42.4 at 3.5x baseline cost.

Why it matters: The two teams break down the cost and payoff of agent harness techniques, comparing multi-agent collaboration, deeper single-model thinking and terminal pipeline changes.

Nov 9Monday

Computing Life · Share · Yage

Kimi and Zhipu align hosted agent APIs with Claude's, so is the protocol race already over?

Kimi Hosted Agents and Zhipu's Z Managed Agents both follow the four-part control plane of Claude Managed Agents, splitting role, environment, session and event. The article reads this as protocol convergence among hosted agents and weighs what it means for picking a vendor.

Why it matters: Kimi and Zhipu adopting Claude's four-part hosted interface gives long-running tasks a client abstraction that carries across vendors.

Nov 8Sunday

Hacker News front page

MaxxBench tests whether DeepSWE's high scores survive altered tasks

MaxxBench changed the requirements or starting code of 89 DeepSWE tasks to see whether coding agents adapt to the changes. It found some models favor guessing at hidden tests. Models that scored close together on the original tasks diverged sharply after the edits.

Why it matters: Small edits to familiar tasks split models that had looked equal, adding requirement adaptability as a way to compare coding ability.

Nov 2Monday

Computing Life · Share · Yage

Why LLM inference splits prefill and decode, and how that leaked user code at Cloudflare

Prefill-decode disaggregation speeds up inference by splitting the two stages, but Cloudflare's GLM-5.3 deployment leaked cache across tenants through a request-handling bug. Its postmortem says 155 long-context replies mixed in other users' conversations or code: over-limit requests were replaced with a one-token placeholder yet still crossed nodes, and decode nodes read stale cache that had not been overwritten.

Why it matters: Cloudflare's fix adds transfer-count checks and delayed memory reclamation, folding data overwrite and page ownership into safety checks when nodes exit abnormally.

Oct 30Friday

The Decoder

Epoch AI: GLM 5.3 builds full Firefox exploit chains in under 48 hours with no human help

Epoch AI found that the open-weight model GLM 5.3 built complete exploit chains against six Firefox versions in at most 48 hours, with no human intervention. It exploited the gap between a patch going public and reaching users.

Why it matters: Epoch AI shows how GLM 5.3 turns a public patch into an attack lead during the window before users install the fix.

Hacker News front page

Study finds GLM 5.3 can break Firefox by exploiting the patch time gap

An Epoch AI experiment found that the open-weight model GLM 5.3 can break six recent versions of Firefox by exploiting the gap between a patch landing and users updating. It reads and exfiltrates local files with no further user interaction.

Why it matters: The experiment points at the window between a fix being published and the update reaching users as the exposed surface.

Oct 17Saturday

Hacker News front page

Neuralwatt on how agent harnesses affect energy use and token efficiency

Neuralwatt analyzed cloud traffic and experimental requests and found that an agent harness editing an existing prompt breaks the cache, adding recomputation and energy use. With the same model and similar input and output lengths, requests that mostly miss the cache use about twice the energy of cache-hit requests, and about 2.5 times for GLM-5.3. When context passes 128k tokens, one broad re-read costs prefill compute roughly equal to 600 ordinary turns.

Why it matters: Neuralwatt traces cache misses request by request and frames stable prompt prefixes and appended state changes as ways to cut an agent's repeated computation.

Oct 7Wednesday

Hacker News front page

Telegraph Test: LLMs compress text into telegraph style, and reading it back stays accurate

Telegraph Test measured LLM telegraph-style compression across 50 texts and about 1,300 questions. Accuracy when another model reads the compressed text ran 0.99 to 1.10 times the plaintext baseline. With lowercase instructions, GLM-5.3-Flash cut tokens by 48.4% but answered in telegraph style at an accuracy ratio of 0.81. gpt-5-mini, which cannot turn off reasoning, cost about twice as much to write the compressed record as plaintext.

Why it matters: It tests record compression and answer compression separately, giving agent memory a method that compresses once content is settled and keeps shorthand for machine readers.

Sep 30Wednesday

The Decoder

Anthropic says Zhipu's open-source GLM-5.3 nears Claude Mythos Preview on exploit capability

Anthropic published an evaluation saying Zhipu's open-source GLM-5.3 comes close to its Claude Mythos Preview on exploit capability. The comparison puts a specific number on how far open models have closed the gap with frontier systems on offensive cyber tasks.

Why it matters: Anthropic and CAISI test data show the measured gap between open models and frontier models on cyberattack capability, plus concrete paths where safeguards were bypassed.

Sep 29Tuesday

AI HOT (Curated Pool)

Anthropic evaluates GLM-5.3: builds end-to-end exploits, safeguards easily bypassed

Zhipu AI's GLM-5.3 is the first open-weight model to approach Claude Mythos Preview in autonomous exploit development. On ExploitBench targeting Chrome V8, it succeeded in 50 of 410 attempts (12%) versus Mythos Preview's 56 (14%). On Anthropic's internal binary exploitation benchmark, GLM-5.3 achieved 4% full control-flow hijack success; Mythos Preview hit 6%. Previous models like Claude Opus 4.6 and GLM-5.2 scored zero on both. The bigger concern is safeguards: simple jailbreaks push GLM-5.3's compliance with malicious orders from 0% to 64% with a false cover story, 92% with prefilled reasoning, and 100% when abliterated. Claude models stayed at 0% across the same tests. NIST's CAISI independently called GLM-5.3 "the most cyber-capable open-weight model released to date," lagging the US frontier by about four months. The post does not disclose GLM-5.3's parameter count, training data, or release format details.

Why it matters: Anthropic red-teamed Zhipu GLM-5.3 for autonomous exploit capability: 50 successes in 410 ExploitBench attempts, close to Claude Mythos Preview's 56, while several other models sat near 0%.

Sep 22Tuesday

Computing Life · Share · Yage

Three AI Coding Stories, Three Numbers to Read

ZCode was caught silently packaging entire Git histories for cloud upload, with .git objects making up 86.6% of snapshots. HarnessTax benchmarked Claude Fable 5 across frameworks: Claude Code and minimal Pi achieved near-identical success rates but a 2x cost gap. A Microsoft architect replaced multi-model agent loops with a single model reading skill docs and calling tools directly—halving API calls but increasing total tokens by 22%. Each story unpacks one number and a reminder to check which layer a metric actually measures.

Why it matters: A single sample shows ZCode packaging the whole workspace, .git included, which can carry secrets and internal branch names to the cloud.

Sep 20Sunday

Computing Life · Share · Yage

Feedback Engineering: Where Agent Automation Gets Stuck, and for How Long

Z.ai published a postmortem on using a GLM-5.3-driven Infra Agent to deploy inference on a domestic chip cluster. The key insight: giving an agent only an end-to-end score traps it in blind guesswork. Splitting verification from diagnosis—with layered, fast, localizable feedback—lets the agent trace issues to specific code paths. Three real cases (precision loss, GIL contention, redundant kernel compute) show how diff comparisons, timeline traces, and micro-benchmarks guide root-cause analysis. End-to-end throughput reached ~3× baseline, but the vendor notes this combines multiple techniques and lacks an ablation study without diagnostic feedback. The engineer's role shifts to designing feedback environments, setting boundaries, and reviewing high-risk changes.

Why it matters: Z.ai reviews deploying an Infra Agent on a domestic chip cluster, proposes feedback engineering that separates acceptance from diagnosis, and walks through precision, concurrency and operator failures.

Sep 18Friday

Hacker News front page

ZCode coding agent silently uploads your entire Git history; only Z.ai holds the decryption key

Developer ferstar reverse-engineered ZCode, Z.ai's desktop coding agent, and found it silently packs the entire workspace—.git history, LFS cache, reflogs, global configs—encrypts it, and uploads to Aliyun OSS whenever logged in. A 345MB commercial workspace became a 313MB encrypted archive; .git alone was 86.6%. The app uses envelope encryption: the symmetric key is wrapped with an RSA public key delivered by Z.ai's server, and the private key lives only in Z.ai's cloud. The user cannot decrypt their own data. The upload pipeline was reconstructed from the client's app.asar: request credentials from zcode.z.ai, pack and encrypt locally, POST directly to Aliyun OSS. In-app privacy toggles don't stop it, and the privacy policy doesn't mention it. The post hit 276K views; a Chinese-language alert urged users to disable ZCode. If you run GLM locally, remember: open weights don't make the closed harness safe. The only working defense is keeping projects outside ZCode's reach or not using it.

Why it matters: Developer ferstar's reverse engineering found that after login ZCode encrypts and uploads the whole workspace, including full .git history, to Alibaba Cloud OSS, with the decryption key held only by Z.ai.

Hacker News front page

ZCode silently packages your entire Git history, encrypts it, and uploads it to Alibaba Cloud OSS

A user found that Zhipu's AI coding desktop app ZCode, when logged in, packages the entire workspace—including .git history, LFS cache, and reflogs—encrypts it, and uploads it to Alibaba Cloud OSS. The RSA public key is delivered by the server on the fly, and the private key lives only in the cloud, so you can't decrypt the multi-hundred-MB file sitting on your own disk. A 313MB .enc file with 564 failed upload attempts was found in ~/.zcode, showing the client keeps retrying. UI toggles don't stop it, and deleting files doesn't help—the client recreates them. The only working defense is locking the cache directory to read-only. The post does not say whether Zhipu has responded.

Why it matters: The author found a 313MB encrypted snapshot in ~/.zcode, traced nearly all of it to .git history, and showed the UI toggle does not stop the upload.

Sep 17Thursday

Hacker News front page

GLM built its own inference infra on 100k+ Chinese accelerators, tripling throughput in under two weeks

Zhipu AI disclosed how GLM-5.3-Flash inference was built from scratch on a cluster of over 100,000 Chinese-made AI accelerators. The team faced limited chip memory, low bandwidth, and an immature software ecosystem. Instead of relying solely on human engineers, they deployed an Infra Agent powered by GLM-5.3 that turned sparse end-to-end metrics into fine-grained, attributable feedback—kernel-level correctness checks, microbenchmarks, and execution traces—so the agent could pinpoint bottlenecks. Combined with tensor parallelism, W8A8 quantization, mixed-precision KV cache, and an Encode-Prefill-Decode disaggregated architecture, end-to-end throughput improved roughly 3× over the initial baseline, with per-token cost reaching parity with mainstream NVIDIA GPUs. Within a week of launch under the anonymous name Ox-Alpha, the model processed over 62 trillion tokens and became the most-used model on both OpenCode and OpenRouter.

Why it matters: Zhipu built GLM-5.3-Flash's production inference service from scratch on over 100,000 domestic accelerators, with much of the debugging done by a GLM-driven Infra Agent.

Sep 4Friday

TechCrunch · AI

Abliteration.ai turns removing AI guardrails into a service

Abliteration.ai launched a platform hosting open-weight models with safety guardrails removed, including Z.ai's newly released GLM-5.3. Users can query them via browser or API. The company frames it as a tool for red teams and offensive security—if a model refuses to write exploit code, defenders can't reproduce attacks. The same removal also enables misuse; the post doesn't detail what access controls are in place.

Why it matters: Abliteration.ai now sells hosted access to open models with their safety guardrails stripped, through a web interface and an API, so the line between red-teaming and abuse is the thing to read for.

Aug 30Sunday

Computing Life · Share · Yage

The value of multimodal models isn't understanding images—it's deciding to look

Meta, Z.ai, and DeepSeek each released multimodal models in August with strikingly similar demos: the model observes a video or screenshot, calls tools to generate a webpage, slides, or a mini-game, then inspects its own output. This shifts vision from a passive input channel to an action the model initiates. The article likens it to the 2023 shift from static RAG to agentic RAG, but notes the loop direction is reversed—here the model self-verifies after producing. Evaluation moves beyond image Q&A: Meta's WildArtifactBench uses pairwise comparisons and Elo scores to assess full artifact creation. Training also changes; both GLM and Meta train models in generate-inspect-revise loops, logging interaction trajectories as training data. For builders, the key question is no longer static image accuracy but whether the model can complete an observe-generate-inspect closed loop.

Why it matters: Three labs' multimodal demos share one action pattern: watch, call a tool, then review the output, giving a common thread for comparing multimodal ability.

Aug 29Saturday

AI HOT (Curated Pool)

Zhipu open-sources GLM-5.3 weights, targeting agentic coding and cyber defense

Zhipu released GLM-5.3 weights for local deployment and commercial use. It scores 60 on the AA Intelligence Index, matching closed-source flagships like Claude Fable 5 and GPT-5.6 Sol, and ties with Kimi K3 for top open-source model. The model excels at complex coding, cybersecurity, and long-horizon tasks. Zhipu added two extra weeks of safety review before release due to its advanced cyber capabilities. Organizations with over $10B annual revenue need a security audit before offering it as an external model service.

Why it matters: Zhipu open-sourced GLM-5.3 weights and allows commercial local deployment, and it ties Kimi K3 for first among open models on the AA intelligence index at 60, close to closed-source flagships.

Aug 26Wednesday

TechCrunch · AI

Z.ai confirms it built Ox Alpha, the anonymous model topping leaderboards

Z.ai confirmed it is the lab behind Ox Alpha, the open-weight model that appeared anonymously on OpenRouter and immediately topped rankings. The company calls it the newest GLM iteration, built for coding, sustained agentic work, and multimodal reasoning. Weights drop Wednesday for developers to build on. Earlier GLM-5.3 already matched Anthropic's Fable 5 on some benchmarks. Ox Alpha adds more pressure on frontier pricing from OpenAI and Anthropic.

Why it matters: Z.ai confirms it built the anonymous model that topped OpenRouter and plans to release weights.