AI Coding Live News

ggml-org released llama.cpp v0.6.0 on Oct 5 (v0.5.0 was Sep 23). It brings GLM-5.3-Flash (GLM5-Next, a 320B text+vision hybrid MoE) into a stable tag, adds MTP speculative decoding for Qwen4Exp (~1.5x decode on DGX Spark), introduces the llama_batch_ext / llama_process() API for mixed token and embedding batches, adds few-row MMA mat-mul kernels on Metal (up to ~3x for speculative and batched decoding), and bumps ggml to v0.26.0. Session file formats are bumped too.

Cantina Security and Yeta Labs released apex-flash-1 (MIT), their first open-weights security research model: a GRPO post-train of GLM-5.3-Flash meant to work as a focused worker under a larger orchestrating agent, reading code, using tools, building exploits and verifying them against a running target. On 60 tasks from 20 held-out vulnerability cases it scores 66.7% pass@1 (40/60) vs 60.0% for base GLM-5.3-Flash and 71.7% for Claude Opus 5 High, at an estimated $2.38 per run vs $74.68 for Opus 5 High. An experimental abliterated variant ships alongside.

MiaAI Lab (@MiaAI_lab) added a TP=3 path to GLM-5.3-Flash EXL3 for three DGX Sparks: +28% decode and +9% prefill vs TP=2, with 3.2M KV cache and 1M context, plus optional NFS weight sharing from the head node (NFS_SHARE=1). Repo: github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks.

GLM ecosystem account Lou (@louszbd) said all Coding Plan users get higher GLM-5.3-Flash usage: unlimited in ZCode from 8:00–18:00 PT daily, and — after feedback on Hermes, OpenClaw, and similar coding agents — double the standard Flash quota in other agents, applied automatically. It aligns with Zhipu lead Zixuan Li’s Sep 3–20 promo window and is a real quota bump for subscribers running GLM Flash coding agents.

Unsloth says optimized local GGUF decoding makes GLM-5.3-Flash up to ~3.3× faster (typically ~1.6–3.4×), with multi-token prediction (MTP) and faster long-context decoding. 3-bit quants still target ~128GB setups via Unsloth Desktop or llama.cpp; docs and Hugging Face GGUF weights are updated.

SGLang says it is working with @zai_org and partners to harden production serving for the GLM-5.3 family. After community and Cloudflare reports, chat-template tool-result reordering now early-exits instead of scanning every block; GLM-5.3 and GLM-5.3-Flash templates are updated and live—material for tool-calling coding-agent stacks.

Z.ai released GLM-5.3-Flash, a native-multimodal 320B-A18B MoE with a 1M-token context, and confirmed it was Ox Alpha—the anonymous model that went viral on OpenRouter. Weights are MIT-licensed, the model ran entirely on Chinese AI chips, API input is /bin/sh.15 per million tokens, and it matches Claude Opus 4.8 on Z.ai’s real-world coding bench.

Z.ai launched GLM-5.3-Flash, a natively multimodal 320B-A18B MIT-licensed model with a 1M-token context, previously previewed as the stealth Ox Alpha. On Z.ai's Code Bench it matches Claude Opus 4.8 at $0.15/$0.50 per million input/output tokens, and it runs entirely on Chinese AI chips.

Ollama said GLM-5.3 and GLM-5.3 Flash (formerly Ox Alpha) are fully live on its cloud, hosted in the US and Europe with zero data retention. First-class launches include ollama launch claude --model glm-5.3:cloud and ollama launch opencode --model glm-5.3-flash:cloud, plus Claude Desktop and a raw API. That moves past a Flash-only preview: the full coding model now drops into the same agent harnesses developers already use, so teams can pick strength versus cost without leaving Ollama’s private endpoints.

Unsloth released 3-bit GGUF quantization for Z.ai's GLM-5.3-Flash (Ox Alpha), enabling local on-device inference for the 320B-A18B multimodal MoE on 128GB RAM workstations without datacenter GPUs.