Guides
Guides
Practical, plain-English guides to model compression, MLX on Apple Silicon, and AI governance, each linking to the research and products behind it.
Content governance guardrails for enterprise AI
Content governance sets the rules for what enterprise AI may produce and who approves it; LLM guardrails enforce those rules at runtime.
Content governance guardrails for enterprise AI
Content governance sets the rules for what enterprise AI may produce and who approves it; LLM guardrails enforce those rules at runtime.
Shepherd: deploying Bounded AI, explained
Shepherd is the deployment pipeline of the Bounded AI Architecture (BAA): it compresses a frontier model into a Ram reasoning model, certifies it with…
AI workforce enablement: turning staff into AI-capable teams
Workforce enablement gives people the competency, authority, and mindset to run AI responsibly, grounded in our RAI and AI Innovation frameworks.
4-bit quantization for MoE models in Apple MLX
Why 4-bit quantization for mixture-of-experts models differs from dense models: router sensitivity, per-expert bit allocation, the long tail of low-activation…
Gemma 4 on MMLU-Pro, explained
What MMLU-Pro measures, what Gemma 4 31B scores (85.2%), and why a RAM-compressed 31 GB build matched that full-precision baseline.
Sharding big models in MLX: the no_gpu_multi(split) path
What model sharding and split placement mean on Apple Silicon unified memory: streaming shards to fit trillion-parameter models on Macs, Metal buffer limits…
MLX quantization on Apple Silicon: a practical guide
A hands-on guide to quantizing local LLMs with mlx_lm.convert on Apple Silicon: bit widths, group size, the dtype footgun, and the pitfalls the official MLX…
MLX quantized model memory footprint and tokens per second on an M4 Max
How to predict a quantized model's RAM footprint from bits-per-weight, why decode speed on an M4 Max is memory-bandwidth-bound, and what actually fits…
Fixing the mlx-vlm "batched vision path does not support KV cache quantization yet" error
What triggers the mlx-vlm ValueError about the batched vision path not supporting KV cache quantization, and four concrete fixes: drop kv_bits, use batch size…
What “expert activated” means in a mixture-of-experts model
A plain-English explanation of MoE routing: top-k gating, softmax scores, why an expert activation is a per-token event, and how softmax confidence scales…
Expert pruning in MoE models: pitfalls and practice
Experts that look dead on a calibration set aren't dead in production. Why activation-frequency pruning is dangerous, what it breaks, and what to measure instead.
NVFP4 and MXFP4 quantization in MLX on Apple Silicon
What separates NVFP4 from MXFP4 (block size, scale type, E2M1 elements), where MLX and Apple Silicon support these 4-bit formats today, and why quality below…
