Nvidia Posts a GLM-5.3-Flash Checkpoint Sized for Two DGX Sparks

The FP4 build fits two 128 GB DGX Sparks with 2M tokens of FP8 KV cache.

@spark_arena flagged a new NVIDIA checkpoint for GLM-5.3-Flash, one that focuses specifically on local AI. The post links nvidia/GLM-5.3-Flash-FP4-Local on Hugging Face and gives one hard spec: the model fits on 2 x NVIDIA DGX Spark 128 GB with 2M tokens of FP8 KV Cache. Two desk boxes, 2M tokens of headroom, no cloud in the picture.

Replies show the local crowd recalibrating. @CodeBuilder_ notes the checkpoint dropped the same day he swapped GLM-5.3-Flash out for Qwen. @tysn_dev is very interested whether the KL beats EXL3 4bpw. Neither question has an answer in the thread, so the checkpoint page is the only spec sheet published so far.

From @spark_arena

Agents & Coding

A Beginner’s Tour of Grok Bot, at $100 a Month

Charlie Chang walks through bots with their own computers, MCP tool links, and scheduled routines.

Charlie Chang’s tutorial covers Grok Bot setup, bot creation, and routines for beginners. He is on the Grok Plus plan at $100 per month billed monthly, and notes the tool comes included with eligible Cursor, Super Grok, or Teams plans. Each bot gets its own screen inside one shared Linux computer. He connects Gmail, Slack, QuickBooks, and Notion, then sets a 9 a.m. routine that turns Slack and email into a morning report.

From youtube.com

Matt Pocock’s Agent Habit Is Asking for State Machines

He says every reducer an agent writes makes future updates trivial and testable.

@mattpocockuk says he is addicted to getting agents to build state machines. Every time an agent produces one, or even a plain reducer, he feels a huge bump in sturdiness, and he knows future updates will be trivial and testable. His recommendation is aimed at frontend work.

From @mattpocockuk

One STYLE.md Makes Claude Write in Your Voice

Zephyr built the file from his own emails and rambles; drafts now sound like him talking.

@Zephyr_hg’s method is one short STYLE.md that Claude reads before it writes anything. He built his from his own rambles and emails, and reports the drafts now sound like him talking. For people running things alone, he argues every outbound message has to sound like you, and the post links his full setup and the prompt.

From @Zephyr_hg

Models & Hardware

Which Model Made Local Feel Good? Readers Answer

Qwen3.8-Flash-Next and GLM 5.3 Flash dominate the replies to @sudoingX’s question.

@sudoingX asked which model first made local inference genuinely good. @JoelDeTeves names Qwen-3.5-27B, then calls Qwen3.8-Flash-Next running on under 64 GB VRAM the craziest moment since. @0x00Sector calls Qwen3.8-Flash-Next his battle horse. @Morrowmake picks GLM 5.3 Flash at high speed and says everything before felt like a toy. @SMT_Solvers answers Qwen3.5 and adds that Qwen3.8 is falling behind his needs.

From @sudoingX

Sources

Thursday, October 15, 2026 Back to the Archive Saturday, October 17, 2026