Topic hub
LLMs & Models
Understand large language models, fine-tuning, prompt engineering, and model selection.
All posts · 41
LLMS & MODELS01
How turbo-fieldfare Runs Gemma 4 26B in Just 2GB RAMturbo-fieldfare runs Gemma 4 26B-A4B on an 8GB M2 Mac at 5-6 tokens per second by streaming MoE experts from SSD instead of loading the full model into RAM.
LLMS & MODELS02
Self-Host Kimi K3: GPU Sizing for 1,561 GB of WeightsKimi K3's checkpoint is 1,560.9 GB, not the 594 GB or 1.4 TB every guide printed. 8x H200 cannot load it. The shard-sum command and the vLLM flags.
LLMS & MODELS03
LLM Latency Targets: TTFT and Tokens/Sec SLOs in 2026LLM latency, measured: TTFT and tokens/sec for 30+ model configs, the P50/P95 targets to set by workload, and the one numeric target any provider publishes.
LLMS & MODELS04
DeepSeek V4 and GLM 5.1 Hosting: 4x Price Spread in 2026The same DeepSeek V4 Pro weights cost $0.87 to $3.48 per million output tokens depending on who serves them. Full provider table and crossover math.
LLMS & MODELS05
vLLM KV Cache OOM: The Flags That Actually Fix Itgpu-memory-utilization now defaults to 0.92, and CUDA graph profiling changed what it buys you. How to read vLLM's memory log and stop guessing at OOM.
LLMS & MODELS06
Kimi K3 vs GLM-5.2 vs DeepSeek V4 Pro: Which to RunKimi K3 took #1 on Arena Frontend Code at 1,679 points. Weights shipped July 27 at 1,561 GB, so compare it to DeepSeek V4 Pro and GLM-5.2 on footprint.
LLMS & MODELS07
Prefill-Decode Disaggregation for LLM Serving at ScalePrefill-decode disaggregation splits LLM inference across separate GPU pools to smooth P99 latency. DeepSeek serves 608B input tokens a day this way.
LLMS & MODELS08
Reasoning Models Burn 3-5x More Tokens: OckBench 2026Reasoning models can cost 5.1x more tokens than a flagship for the same accuracy. OckBench 2026 quantifies the overthinking tax and how to cap it.
LLMS & MODELS09
Run GLM-5.2 Locally: Hardware, Quantization & vLLMRun GLM-5.2 locally: 744B MoE, 1M context, MIT license. Unsloth's 2-bit GGUF squeezes 1.51 TB down to 239 GB at 3-9 tok/s on a Mac Studio.Explore other topics