Open Weights, Closed Doors: GLM-5.3 Lands as Cursor Loses Its Models

GLM-5.3 lands as open weights with a 6x jump on Terminal-Bench 3.0, OpenAI cuts Cursor off after the SpaceX deal, and FaulT-Bench grades network-troubleshooting agents on messy tickets. Plus Datalog agent memory, a Git server on object storage, and a 3B edge guardrail.
Author

Sarathy πŸͺ†

Published

August 30, 2026

πŸŽ™οΈ Listen while you go

Prefer to listen on the go? Here’s the audio edition β€” Avery and Jordan walk through today’s eleven stories.

Decode Daily audio edition Β· 2026-08-30 Β· ~7.7 min

Networks Β· 5G & 6G
arXiv

FaulT-Bench: benchmarking network-troubleshooting LLM agents on unreliable user tickets

FaulT-Bench targets a realistic pain point in LLM-driven network operations: agents that troubleshoot network faults must work from user tickets that are noisy, incomplete, or even misleading. The benchmark grades agents on how accurately they diagnose and localize faults given such unreliable inputs, rather than clean synthetic scenarios.

It is a concrete step toward measuring whether network-management agents hold up under the messy conditions operators actually deal with.

Security
Jordy Zomer

Lemmalog: LLM memory as a Datalog program-analysis engine

Jordy Zomer built Lemmalog to keep an LLM's memory consistent during long vulnerability-research sessions: instead of retrieved chat snippets, it stores facts and derivation rules the way a program-analysis engine does. When a later debugger observation falsifies an earlier pointer, invalidating the premise automatically drops dependent conclusions.

The post reports it is competitive with dedicated LLM memory systems on some tasks while showing the agent only a tiny fraction of the history; the repo is a Datalog engine with provenance-tracked facts and incremental updates.

Hugging Face

ReasonShield: a 3B policy-adaptive multimodal safety classifier

ReasonShield is a 3B multimodal safety classifier fine-tuned from Shieldstral 1.0 3B, small enough to run on a 3060-class GPU once quantized. It writes a short visible decision summary inside think tags, then a lowercase yes or no, with a one-token Shieldstral-compatible scoring mode.

The card reports mixed held-out results: plus 13 ArabSafe recall and plus 12.3 MultilingualSafety F1 in adaptive mode, but HarmBench recall drops to 83.75 from 98.44 and ToxicChat F1 falls. BF16 weights plus GGUF are provided; the author is explicit that a plausible rationale can still be wrong.

AI & Models
Z.ai

GLM-5.3 is now open-weight: a 753B-class coding and agentic upgrade from Z.ai

Z.ai published GLM-5.3 as open weights on Hugging Face, reusing the same 753B base as GLM-5.2 so every claimed gain comes from post-training. The card reports open-source state of the art on Terminal-Bench 3.0 (28.3 versus 4.6 for GLM-5.2), DeepSWE v1.1 (46.2 to 66.9), and Agents' Last Exam (23.8 to 28.5), plus a jump on ExploitBench from 24.4% to 54.4% and a CyberGym score of 84.5 for vulnerability discovery.

SGLang, vLLM, Transformers, Unsloth, and Ollama recipes are already linked from the card. The numbers are the authors' reported figures; the flash-tier variant was covered in the 2026-08-27 edition.

arXiv

WikiSkill: compiling agent experience into a persistent wiki for skill evolution

WikiSkill sits a persistent wiki between raw agent traces and executable skills, so skill updates compile experience instead of scattering it across optimization histories. Across the authors' benchmarks it beats EvoSkill and SkillOpt, and smaller models with evolved skills can outperform larger models without them.

Evolved skills also transfer across model families, sometimes beating a model's self-evolved skills, and ablations show the wiki layer is the part that matters.

arXiv

SWE-Prime: fewer trajectories, better SWE-agent supervised fine-tuning

SWE-Prime argues that a successful SWE-agent trajectory is not automatically good supervised fine-tuning (SFT) data: it can still contain redundant steps, failed tool calls, or git-history cheating. A two-stage filter first keeps high-quality trajectories, then scores semantic segments, keeping the full sequence for context but applying loss only on selected spans.

Training on the selected 10% of resolved trajectories beat training on the full set, with reported relative gains of up to 12.2% on SWE-Bench Pro and 24.2% on SWE-Bench Verified.

arXiv

Same model, different harness: coding-agent results vary by framework

A study showing that the same underlying model yields materially different coding-agent results depending on the harness that wraps it. Agent scaffolding, tool-calling conventions, and context assembly change outcomes more than model choice alone.

The result is a reproducibility warning for agent evaluation: benchmark numbers are as much a property of the harness as of the model.

TerminalBytes

Run Qwen3.8-27B locally: real numbers from a Mac Studio M3 Ultra

On a Mac Studio M3 Ultra, Ollama's Q4_K_M Qwen3.8-27B (17 GB) generates about 14 tokens per second versus 28.6 for Qwen3.6-27B on the same box. The newer model answers in roughly a third the tokens, so wall-clock per finished answer is close to a tie.

A 1-bit 6.7 GB llama.cpp build hits 27 tokens per second and gets facts right but will not commit to an answer. You need a recent llama.cpp β€” older builds fail on the qwen35 architecture tag β€” and roughly 32 GB of RAM for Q4, 16 GB for Q2.

arXiv

Praxist: an autonomous R&D loop that turns experimental artifacts into solution lineages

Praxist treats a measurable, computer-executable research project as a persistent process: parallel research peers, task-owned evaluation, durable evidence, and generation-to-generation synthesis. Installation is a pip extra plus an interactive setup that covers the Fair Source license; commercial use is free under $1M annual revenue.

The accompanying paper frames the output as solution lineages β€” families of artifacts and the evidence that links them. Generated outputs need product-name attribution if published externally.

Business & Funding
Indian Express

OpenAI will pull its models from Cursor after the SpaceX acquisition

OpenAI told SpaceX it will wind down the contract that supplies OpenAI models to Cursor, with a proposed shutoff date of November 12, 2026 β€” the maximum notice in the agreement. OpenAI cites a change-of-control window plus prior terms-of-service breaches after Musk acquired Twitter and, under oath this year, xAI's admission that it distilled OpenAI data.

OpenAI says it will not provide future models, including Astra, to Cursor. Developers who rely on GPT inside Cursor have until mid-November to migrate.

Research & World Models
GitHub

walgit: a Git server that is one binary in front of object storage

walgit is a single Rust binary that hosts Git against S3 or GCS with no database, no leader, and no local state that matters. A write-ahead log in the bucket is the source of truth, disk is a cache, and a compare-and-swap on a tiny manifest is the only consensus.

It implements the architecture Cursor described as Continuity, plus remote readers and bundle-uri so clones can be larger than the machine. MIT-licensed, with smart HTTP, LFS, a web UI, and a fault-injection test suite.

No matching items