Jon

Pushing for local AI sovereignty among devs
4x spark enthusiast

[Jon @jonathan_conn]

The new meta for 4x spark setups

Glm 5.2 as your orchestrator
DSV4F as your worker

Hard to beat this combo right now

[theo @theoluu44]

oh Kevin Durant? you mean the 240x open-source AI angel investor?

[Jon @jonathan_conn]

@TechMDAI 100% the biggest problem we have with glm 5.2 is the speed

It's by far our favorite model but we can only run it around 25-35 tps on the sparks

A glm 5.3 flash would be a complete game changer

[Jon @jonathan_conn]

Rumors are glm 5.3 flash

[OpenCode @opencode]

after 56 straight days of deepseek being #1 on our charts it has finally been dethroned

[Jon @jonathan_conn]

@mr_r0b0t @NousResearch Anyplace I can get the skill?

[unusual_whales @unusual_whales]

AT&T is using open source models to decrease the costs of its Anthropic bills, per the Information

[Jon @jonathan_conn]

@SergiiioBS

[Jon @jonathan_conn]

@Tech2Wild Anyone else see it over thinking everything ?

[Jon @jonathan_conn]

@landontgreen @NVIDIAAI Congrats! 2 is the magic number the spark starts to fly

[Jon @jonathan_conn]

@TechMDAI I used naddod had a great experience

[Jon @jonathan_conn]

@loktar00 Thanks for tracking this

[Macro Charts @MacroCharts]

They're curing cancer, and you're bearish?

[Lotto @LottoLabs]

How cloud bros see me and my 3090s running cracked qwen 27b 3.8 weights

[Jon @jonathan_conn]

Short film inspired by my cat
Used seedance 2.5 + lyria

Total cost $35

[SergeB @SergiiioBS]

3.565 downloads in the 27B model checkpoint 🤯

[Jon @jonathan_conn]

@StefanMaier H3 native, no add ons, split h3 layers between the two gpus

[Jon @jonathan_conn]

Generated with minimax H3 locally on 2x3090s and comfyui

Pretty impressive how far local video has come

[SGLang @sgl_project]

SGLang is the first engine to develop and land Breakable CUDA Graph (BCG), the full CUDA Graph, and graph memory reuse.

- BCG drops torch.compile for faster setup and broader compatibility
- Full graph capture brings prefill latency down on dynamic workloads
- Memory reuse keeps the graph footprint fixed as coverage grows.

Read the design details in the blog 👇

[Jon @jonathan_conn]

Opening github on monday to see the stupid fucking outage unicorn staring back at me

[Hikari∣LocalLLM⚡ @Hikari_07_jp]

Even if we buy a GPU for local LLMs, we still end up burning through tokens on the actual GPU work. That's us:

[Jon @jonathan_conn]

@Snixtp Beautiful, is that 2 6000s and 4 3090s?