My journey to develop AGI spans 25 yrs, including 10+ yrs thinking about technical & societal perspectives at Google DeepMind. AGI is on the horizon - we need deeper understanding of its implications. To help, we've created the DeepMind Institute.

I'm 38. Solo founder from Germany. Looking to connect with more builders & indie hackers!

my model recommendations for every gpu class, 6gb card to a 256gb double dgx spark cluster, from my bench book: ────── 6GB ────── > bonsai 27b at 1-bit Q1_0: 20.5 tok/s, 8K context, 3.5GB of weights. a 27b thinking on a 6gb card is the wildest small vram result i own ────── 8GB ────── > gemma 4 E4B Q4_K_M: 42 tok/s and 656K context fits, sliding window attention stretches 8gb an order of magnitude past the qwens > bonsai 27b Q1_0 on an ampere 8gb: 42 tok/s fresh, 128K context, the smallest card that runs a full 27b agentic loop ────── 12GB ────── > qwen 3.5 9b Q4_K_M: 50 tok/s and dead flat from 4K to 512K context, 43 under real agent load > qwen 3.5 9b Q8_0: fits with 64K context when output quality matters more than ceiling, the quality pick > gemma 4 12B at Q4: sliding window attention stretches the context budget, the long-context alternative ────── 16GB ────── > qwen 3.8 27b dense squeezed to Q3 class: the 27b fits and flies, 59.5 tok/s on a 5060 ti, 101.3 tok/s on a 5080 with the mtp flag, 40 tok/s even on a 4060 ti ────── 24GB ────── > qwen 3.8 27b dense Q4_K_M: 41 baseline, 65.8 tok/s with the mtp flag, full 262K context resident in 22GB on q4_0 kv. the best model you can put on 24gb > the speed build of the same model, IQ3 class: 43.7 → 75.9 tok/s with mtp, serving in 16.2GB, quality holds on the AD calibration > qwen 3.6 35B-A3B at IQ4: 180 tok/s on a 4090, the moe rocket for this tier > bonsai 27b at 1-bit: 67.9 tok/s with a 786K context ceiling, the long-context freak ────── 32GB ────── > qwen 3.6 35B-A3B Q4_K_M: 236 tok/s on a 5090, 180 on a 24gb 4090, the newest moe pick > qwen 3.5 35B-A3B at Q4: 150+ tok/s on a 5090, pushed to 230 on vllm with fp8 kv, the proven speed monster > qwen 3.8 27b dense at Q6 with the flag: 130 tok/s at 128K, the dense quality pick and nobody has published the 35B at Q6 on a 5090 yet, it fits with headroom, that row is waiting for a name ────── 128GB DGX SPARK ────── > qwen 3.6 35B-A3B Q8: 58.6 tok/s, the fastest general model on one spark > ling 3.0 flash official int4 with mtp: 40.9 tok/s on code, the sprint king under 30K context > ling gguf at Q5_K_M: ~40 tok/s and flat at depth with 256K in memory, the ling marathon quant > laguna s 2.1 nvfp4 with dflash: 45.4 tok/s sustained on code, flat at every depth, the agent workhorse > deepseek v4 flash, 284B at 3-bit: 16.5 tok/s, the biggest brain one box holds, and the reason box two exists ────── 128GB STRIX HALO ────── > qwen 3.6 35B-A3B Q8_0 on vulkan: 53.6 tok/s, 35 at 131K deep, 4 million tokens of context loaded. the daily driver > qwen 30B-A3B class at Q4: ~100 tok/s, the speed pick > nex-n2-pro 397B at 1-bit: 20.5 tok/s with a 1M context ceiling, the stunt that works and run vulkan not rocm on this box, 18% free speed ────── 256GB, 2x DGX SPARK ────── > deepseek v4 flash official fp8, tensor parallel over roce with the dspark drafter: 64.7 tok/s single stream, 145 aggregate, a 1M token window > the new nvfp4 kv path: 67 tok/s mean, peaks at 84 tok/s, 197 tok/s aggregate at 6 streams, 261 tok/s at 16 and decode stays flat from 600 tokens to 143K deep, the depth curve simply does not exist at this scale if your card class is missing a row or you can beat one of these, my repos take pull requests. run it, measure it, own it.