Executive Summary
GLM-5.2 has emerged as the dominant story of the day, widely hailed as the first open-weights model that functions as a genuine daily driver at the frontier level (Opus 7/8 tier, MIT license). Meanwhile, the local AI ecosystem is rapidly maturing around inference optimization, with multiple voices emphasizing that software stacks and kernels matter as much as hardware specs. Codex (OpenAI) is gaining significant traction among developers as an agentic coding tool, and Hermes Agent passed 1,500 contributors. A growing Python-to-Rust migration is underway, with AI-assisted porting showing tangible container size and performance gains.
Key Events
-
GLM-5.2 released to widespread acclaim — Multiple developers and organizations (including Ollama, tinygrad, and independent testers) report it as the first open model that passes the "daily driver" bar, ranked #3 on Agent Arena, competitive with Opus 7/8. Available under MIT license. → link
-
GLM-5.2 generating custom CUDA kernels for DGX Spark — User reports GLM-5.2 is producing optimized CUDA kernels to run DeepSeek-v4 at full speed locally without quantization or pruning. → link
-
GLM-5.2 running at 120 tok/s on networked Blackwell Tinyboxes — Tinygrad confirms the benchmark, positioning $150k dual-tinybox setup as a cloud-free frontier inference solution. → link
-
NVFP4 quantization of GLM-5.2 uploaded — 467GB variant fits on 4× DGX Sparks (~$20k total), making frontier-scale inference accessible at small-cluster prices. → link
-
Hermes Agent hits 1,500 contributors — Nous Research's open agent framework continues rapid community growth. → link
-
Codex gaining developer mindshare — Multiple prominent developers report switching to OpenAI's Codex for daily coding work, citing GPT-5.5 performance, generous limits, and multi-platform support. → link
-
Python-to-Rust port completed with AI assistance — 300k-line Python server ported to Rust via Codex in ~24 hours; image size dropped from 2.2GB to 716MB. → link
-
Local AI inference guide released free — Comprehensive reference for running LLMs locally covering all hardware tiers and software stacks (llama.cpp, MLX, ExLlamaV3, vLLM, SGLang, TensorRT-LLM, Dynamo). → link
-
StepFun Step 3.7 Flash recommended as best DGX Spark model — 198B MoE VLM running at Q4_K_S on single 128GB DGX Spark with full 262K context at ~25 tok/s. → link
-
REAP'd DeepSeek V4 Flash pruned to 82GB for DGX Spark — Custom pruning removes redundant MoE experts, freeing 30GB for working context versus the original 112GB that left only 6GB. → link
Analysis
Patterns: The conversation has decisively shifted from "which model is best?" to "how do I run frontier models locally and efficiently?" The twin narratives of GLM-5.2's quality and local inference optimization are converging — open models are now good enough that the bottleneck is deployment, not capability. Inference engines (vLLM, SGLang, ExLlamaV3) and kernel optimization are being discussed with the same intensity previously reserved for model benchmarks.
Escalation trend: Local AI advocacy is escalating from niche hobbyism to serious infrastructure posture. Multiple voices are framing data sovereignty — law firms, clinics, financial firms — as the primary justification, not cost savings. Simultaneously, the tooling layer (Hermes Agent, Codex, multi-agent tmux orchestration) is professionalizing rapidly.
What to watch next: GLM-5.2's real-world staying power after the initial hype cycle; whether the Rust migration pattern accelerates across other AI infrastructure; DeepSeek V4's local deployment viability as more optimized quants and pruning approaches emerge; and whether Tenstorrent's fully OSS stack gains traction as a hardware alternative.
Tweet Feed
GLM-5.2 / Open Model Releases
@cooltechtipz · 2026-06-21T16:53
Big week for open-source models. GLM-5.2 is Opus 7/8 level and outperforms Gemini. → tweet link
@cooltechtipz · 2026-06-21T11:20
The notable part of GLM-5.2 isn't the 1M-token context alone. It's being able to use all that information effectively while keeping coding workflows fast. → tweet link
@Ex0byt · 2026-06-21T18:47
The race is on 🤗 → tweet link
@Ex0byt · 2026-06-21T14:26
5 days in with GLM-5.2 since I first told you all about its capabilities (and you listened). 5.2 is now generating custom CUDA kernels for DGX Spark to run DeepSeek-v4 at full speed with no quantization or pruning. This directly translates to optimized local serving of GLM-5.2 → tweet link
@tinygrad · 2026-06-20T22:35
The new personal computer revolution is just beginning. → tweet link
@tinygrad · 2026-06-20T22:30
I have on good authority that GLM 5.2 is running at 120 tok/s across two networked Blackwell tinyboxes. $150k and that setup can be yours, either 2x tinybox or 1x tinybox pro. Never pay the cloud again. → tweet link
@TheAhmadOsman · 2026-06-20T21:00
Luke Alonso has uploaded an NVFP4 of GLM 5.2. 467GB, would fit on 4x DGX Sparks (~$20k) → tweet link
@TheAhmadOsman · 2026-06-21T15:59
Qwen 3.5 27B and GLM 5.2 canceled that permanent underclass semi-joke threat btw → tweet link
@TheAhmadOsman · 2026-06-21T14:53
Opensource AI is nothing without its people ❤️ → tweet link
@TheAhmadOsman · 2026-06-21T12:12
What a difference 8 months make. Opensource AI is no longer niche → tweet link
@ollama · 2026-06-20T22:29
RT @Resorcinolworks: @ollama thank you for existing ollama → tweet link
@ollama · 2026-06-20T21:07
Let's go open models! ❤️ → tweet link
@ollama · 2026-06-20T20:06
RT @dwlz: GLM-5.2 is really, really good. → tweet link
@ollama · 2026-06-20T20:01
RT @lassejv: glm 5.2 is awesome → tweet link
@ollama · 2026-06-20T19:50
RT @codebrandes: GLM 5.2 is really good. Wow. → tweet link
@ollama · 2026-06-20T19:49
Let's go open models! ❤️ → tweet link
@ollama · 2026-06-21T02:47
Let's go open models! ❤️ → tweet link
@ollama · 2026-06-21T02:14
RT @matthewfurnari: @ItakGol Yeah I agree. The big winner is going to be Ollama: I've offloaded all my supervisory, code review, and ontolo… → tweet link
@gospaceport · 2026-06-21T01:25
RT @0xSero: Rejoice fellow GPU poors GLM-5.2 GGUFs coming soon dynamic 2bit, 3bit, and 4bit — 3bit will fit on 256GB / 2 sparks — 2bit wi… → tweet link
@cooltechtipz · 2026-06-21T10:43
AI progress is no longer just about better models. More value now comes from how models connect with data, tools, memory, and workflows. The model is only one part of the bigger system. → tweet link
@cooltechtipz · 2026-06-21T18:02
People who consistently spot new opportunities created by AI will outperform people who simply know how to use AI. → tweet link
@cooltechtipz · 2026-06-21T08:30
AI is making technology more useful for people who communicate well and solve problems effectively. → tweet link
Inference Optimization & Kernels
@TheAhmadOsman · 2026-06-21T18:55
Why do I focus on Inference Engines/Software Stacks for your hardware? — 2x RTX 3090s: ~14.5 tok/s → ~64 tok/s moving to vLLM w/ TP=2 — RTX PRO 6000: ~32 tok/s → ~110 tok/s moving to Sglang. So: CUDA/2+ GPUs: ExLlamaV3/vLLM/Sglang > llama.cpp — Edge: llama.cpp > Ollama → tweet link
@TheAhmadOsman · 2026-06-21T16:48
You run Kernels, not models. The model is just a graph. The Inference Engine serves as a scheduler, optimizer, and executor. But the actual work? That happens in the Kernels. Same model, same GPU, same VRAM — Wildly different performance. Because one stack is using optimized fused Kernels that understand your hardware. → tweet link
@TheAhmadOsman · 2026-06-21T09:33
Local AI hardware = capacity × bandwidth × software stack. [Detailed hardware bandwidth comparison table for Mac Studio, RTX PRO 6000, RTX 5090, DGX Spark, Strix Halo, Tenstorrent, etc.] Fitting ≠ serving. The only mental model that matters: 1. What must fit? 2. What bandwidth tier do I need? 3. What software stack can actually deliver it? → tweet link
@TheAhmadOsman · 2026-06-21T10:15
Eric, thank you for making the case for Opensource AI so clear → tweet link
@TheAhmadOsman · 2026-06-21T02:56
DROP EVERYTHING — The bible for running LLMs locally is now available online to read for free. Covers what to use on Laptop/edge, Mac-first, Single RTX GPUs, 2-4+ NVIDIA/CUDA GPUs, production serving, long-context/MoE/routing, cluster orchestration. Software: llama.cpp, MLX/MLX-LM, ExLlamaV2, ExLlamaV3, vLLM, SGLang, TensorRT-LLM, NVIDIA Dynamo. → tweet link
@TheAhmadOsman · 2026-06-21T02:46
I like that saying. Cloud is tragedy of commons. → tweet link
@TheAhmadOsman · 2026-06-21T00:42
Can we stop doing hardware cost to token generation calculations on the timeline please? If you haven't noticed, models keep getting better & more efficient, and hardware prices keep going up → tweet link
@TheAhmadOsman · 2026-06-21T01:01
Oh, you'll break even in 6 years — Bruh, some of you did the math a year ago and told me you'd break even in 27 years. AI is not a car lmao → tweet link
Local AI Hardware & Model Deployment
@sudoingX · 2026-06-21T18:03
the local ai craze is absolutely real, but almost nobody's saying the core part that the harness matters as much as the model. i'm running stepfun's step 3.7 flash on dgx spark through hermes agent. plug and play. point it at the local endpoint and it auto detects and just works. the bloated frameworks can't do this. → tweet link
@sudoingX · 2026-06-21T17:50
anon if you have a dgx spark, the best model to run in 2026 is stepfun's step 3.7 flash. a 198B mixture of experts vision language model, ~11B active per token, Q4_K_S at ~104gb, running on a single 128gb dgx spark under hermes agent. one machine. it holds the full 262K context at ~25 tokens a second and takes image input. → tweet link
@sudoingX · 2026-06-21T16:04
a used 3090, 900 to 1200 bucks, runs qwen 3.6 27b dense and does real agentic work. on your desk. offline. yours. i will never get over this, we're living through the cheapest superpower in human history and most people scroll right past it. → tweet link
@sudoingX · 2026-06-21T16:59
hey @0xSero, i was benching your reap'd deepseek v4 flash on dgx spark and hit a wall loading it. the gguf ships the per-block hyperconnection tensors but every llama.cpp fork i've tried wants a global hc_head_base it doesn't have, so it won't load. which fork and branch did you build it with? → tweet link
@sudoingX · 2026-06-21T15:38
i'm hunting for models that actually fit and run well on single dgx spark, and most don't make it. here's the trap almost nobody post about, a model fitting in 128gb is not the same as it being usable. the full deepseek v4 flash is 112gb on this box, technically it fits, but that leaves about 6gb for context. → tweet link
@sudoingX · 2026-06-21T13:17
i see a lot of misconceptions about qwen 3.6 27b dense lately, and almost all of them from people who've never run it. qwen 3.6 27b dense is the king of the single 3090 tier and most of the skeptics have never even met it. → tweet link
@sudoingX · 2026-06-21T06:59
most people in 2026 are still running one model in one chat tab like it's 2024. here's my real setup, 4-6 agents going at any given time, spread across three boxes on my desk, the dgx spark, the strix halo, the 5090 laptop. all in tmux, all reachable from my phone. the skill was never the model, it's the orchestration. → tweet link
Developer Tools: Codex, Hermes Agent & Coding Agents
@gdb · 2026-06-21T18:23
codex for testing every single feature in your app: → tweet link
@MengTo · 2026-06-21T07:15
Don't sleep on Codex. I've been using it since day one. — GPT-5.5 xHigh + full access is beast mode — Computer use + browser use + spawned threads make agent loops incredibly powerful — Picking up tasks from mobile works surprisingly well — Still using Claude for writing (and waiting for Fable 5), and Cursor when I need raw speed. But these days, 95% of my work happens in Codex. → tweet link
@MengTo · 2026-06-21T17:05
When you let your agents run wild → tweet link
@steipete · 2026-06-21T17:09
RT @petergyang: I used to be a die-hard Claude Code user. Codex has won me over because: → GPT-5.5 is excellent → Fast mode + generous li… → tweet link
@steipete · 2026-06-20T23:29
RT @samdenty: There's an OSS implementation of the codex permission flow in Swift here: [link] → tweet link
@Teknium · 2026-06-21T14:32
We just passed 1500 Contributors to Hermes Agent's repo! Thank you to all the contributors and developers! → tweet link
@Teknium · 2026-06-21T13:26
🫡🫡 → tweet link
@Teknium · 2026-06-21T02:06
RT @NousResearch: [Hermes update link] → tweet link
@Teknium · 2026-06-20T20:23
RT @hermes_updates: [Hermes update link] → tweet link
Python → Rust Migration & Software Infrastructure
@jezell · 2026-06-20T21:02
Why did I port this server from python to rust (besides the obvious)? Well, a large number of the components like lance, lance-graph, datafusion, etc. are all rust anyway and the python wrappers just don't expose all the knobs. Python is great for iteration and exploration, but eventually you need to go deeper. → tweet link
@jezell · 2026-06-20T20:55
Some initial rust port stats. About 300k lines of python in this server originally. Python image size: 2.2 GB — Rust port image size: 716 MB. Took about 24 hours for codex to port this server. → tweet link
@jezell · 2026-06-21T07:24
Rust is the new assembly language. → tweet link
@jezell · 2026-06-21T02:42
My Codex usage this month. Glad I'm not paying API rates... → tweet link
@jezell · 2026-06-21T03:03
People who know nothing about development seem to be consistently able to do more with LLMs than a lot of professional developers. It's weird, but I think they just think bigger. → tweet link
@jezell · 2026-06-21T01:14
SlateDB seems to be powering a lot of the cool new things. → tweet link
@jezell · 2026-06-21T01:13
RT @haipingfu: Introducing Crab: a serverless Git remote storage solution for teams working with large files. → tweet link
@jezell · 2026-06-20T22:05
RT @thehypedotnews: z ai with 127 employees surpassed google on agent arena. z ai's glm 5.2 – open weights, mit license – ranked #3. google… → tweet link
Data Privacy & Local AI Advocacy
@sudoingX · 2026-06-21T12:53
the $20 subscription was never really the point anon. here's what actually changes when you run ai locally yourself. your data never leaves the room you're in, so there's no policy to trust, no terms that quietly change next year, nothing for a government to subpoena. you stop renting your intelligence and start owning your tools. → tweet link
@sudoingX · 2026-06-21T12:59
the most valuable thing you produce isn't your work, it's your intellect. every prompt is a window straight into your mind. you are not the customer, you're the raw material. a gpu, the one thing that keeps your own thinking in your own room, runs the model on your hardware, feeds nothing back, that you won't buy. → tweet link
@sudoingX · 2026-06-20T20:15
if you are thinking to run open models and haven't saved the weights locally, do it tonight. the one you can run today isn't guaranteed to be downloadable tomorrow, models get pulled, deprecated, region locked, quietly removed. the copy on your own drive is the only access nobody can take back. → tweet link
@sudoingX · 2026-06-20T19:58
here are a few businesses i would never run a cloud model in. law firms, financial firms, clinics, startups whose moat is the codebase, accountants. for every one of these, the data IS the business. it leaks, the business is over. this is the part of going local that was never about money. → tweet link
@sudoingX · 2026-06-20T20:06
be honest anon, for the work you actually do every day, do you need a frontier model, or is a good open one already enough? → tweet link
@sudoingX · 2026-06-21T16:45
what's your actual monthly ai bill. just wanna see where people really land → tweet link
AI & Software Engineering Philosophy
@sudoingX · 2026-06-21T07:55
if you actually want to be a software engineer, learn to tell signal from noise early, because the tools won't do it for you. every ai tool is built to hand you output, not to make you better. it's a choice: sit and wait for an answer to drop in a text box, or pair-program with the thing and learn why the answer actually works. → tweet link
@sudoingX · 2026-06-21T07:37
i get called out a lot for how long my build is taking, when someone can vibe-code an app in a weekend. i'm not just building. i'm learning to tell good code from bad code. vibe-coding is genuinely great until the day the model can't fix the bug and neither can you. this is the best time in history to learn to code, not the excuse to skip it. → tweet link
@TheAhmadOsman · 2026-06-20T21:37
Seriously, any talented potential hire will sign immediately if you offered them a sign-on bonus of a GB300 / DGX Station on the spot → tweet link
AI-Powered Products
@levelsio · 2026-06-20T19:53
I normally don't like plugging an AI chat into my projects because it seems too easy and basic. I think you should instead rebuild entire projects from the ground up to be AI first. So you can now talk to the hotel assistant and ask it to find hotels with a weightlifting gym, newly built, highly rated. And it controls the site, moves the map, opens hotels for you! → tweet link