I’ve learned my lesson and I’m going all in. DDR5 is now $7200 for 256GB. Prices will probably double in a year. I failed to buy 3 months ago and I now realize it’s an appreciating asset for engineers. Don’t get talked out of it if you’re on the fence and can buy a config today. The more you learn the more value it yields. I’ve burned over 70 billion tokens this month and the frontier models are smart enough to cover my entire SWE workload. I’m investing because 18 months from now I should be able to train models and run inference for that workflow at home forever. I can serve inference across this plus my 2 DGX Sparks. Several years from now the kernel efficiency and model improvements will hit a point where it’ll hill climb on its own and pay for itself many times over but people will not want to pay $60k for hardware. I could be wrong about this and I’ll eat my hat. Screenshot this and tell me I’m wrong in 2029.

@ItsmeAjayKV Honestly I’m amazed I can run Strata qwen 3.8 next coder quant on B580 12GB VRAM (newest part) (gen 3x8 speeds) The rest is 7 years old now 32 GB 3200 DDR4 Asus ROG b450f gaming Ryzen 5 3600x 64GB swap on a 1TB NVMe drive gen 3x4 wd blue ssd I get usable perf.

Calling out all 8 - 16GB vram users running local LLMs. If you have 64gb ram then whatever llm you're running can be replaced by Qwen3.8-flash-next. Thanks to Strata. If you don't know what Strata is, it's a specialized inference engine, tuned to work for 8 - 24GB vram. To know more checkout @coldniko posts, or my old posts on the same.

Another win for Ling-3.1-flash 🔥! Ling-3.1-flash is also really good in generating motion-graphics. Game ad generation, all in code, no external assets, just html, css and svg. This is also better than deepseek-v4.1-flash. This is nuts, we are getting really good flash models now and am all for it. Way to go @AntLingAGI

How abliterated models can get you pwned 👾 We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts. Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain. In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.