AIHow AI Works

Tag: Inference

Qwen3.8-Max’s Open Weights Are Not the Model You Benchmarked

Alibaba opened its 2.4T flagship — but the download is text-only, thinking-required and 262K context, under a licence that is not Apache 2.0.

Gemini 3.8 Flash: Same Price Per Token, 40% More Per Task

Google's Gemini 3.8 Flash shipped September 2, 2026 at $0.75/$3.75 per million. Artificial Analysis measured cost per task rising 40% anyway.

DeepSeek Plans a 160,000-Chip Huawei Cluster in Inner Mongolia

Bloomberg reports DeepSeek will deploy at least 160,000 Huawei Ascend 950DT accelerators in Inner Mongolia — for inference. Training stays on Nvidia.

Apple Builds Macs for AI Clusters as OpenAI Reportedly Stocks Up

Mac revenue hit $10.35B, up ~29%. The Information reports OpenAI bought tens of thousands of Macs for RL, days after Apple pitched new Macs at local AI.

Qwen3.8-Flash-Next Runs Locally in 75 GB of RAM

Alibaba's 125B multimodal MoE is out as Unsloth GGUFs. The 1-bit build fits in about 75 GB of RAM or unified memory, with no GPU VRAM required.

NVIDIA Claims 30x Efficiency Gain for AI Agents on Vera Rubin

NVIDIA reports Vera Rubin NVL72 delivers up to 30x higher throughput per megawatt and 35x lower token cost than GB300 NVL72 on agentic coding workloads.

Mixture of Experts Explained: Why Every 2026 Model Uses It

Mixture of Experts lets a model have trillions of parameters but only use a fraction per token. Here's how it works and why it took over in 2026.

Self-Hosted vs API: When Running Your Own LLM Pays Off

The raw GPU price is a trap. A 2026 breakdown of the real cost of self-hosting an LLM and where the break-even against API pricing actually sits.

From Token to Transistor: What Happens When You Send a Prompt

Follow one prompt through the machine — tokens, matrix multiplications, the memory that everything waits on, and the chips built around a single operation.

Training vs Inference: Two Economies, Two Opposite Bottlenecks

Training is network-bound and paid once. Serving is memory-bound and paid forever. The crossover point, and why the 2024 chip-demand call was backwards.