• 6 min read
TurboQuant & Anvil: Breaking the KV Cache Memory Wall in Local LLM Inference
How Google's TurboQuant Fast Walsh-Hadamard Transform and our Anvil engine deliver 4.6x KV cache compression with +30-50% throughput acceleration.
• 7 min read
Sub-8B Inference Performance on Commodity Silicon: M4 Max vs RTX 4090
Empirical benchmarking of Solace-Sub8B checkpoints in INT4 AWQ, FP8, and GGUF across Apple Silicon and NVIDIA Ada Lovelace GPUs.
• 5 min read
Why Multi-Teacher Consensus Eliminates Mode Collapse in Reasoning SFT
How training on multi-teacher verified consensus scratchpads prevents student models from memorizing single-model hallucination patterns.
• 6 min read
Quantization Without Degradation: Preserving Multi-Step Reasoning at INT4 & FP8
A practical study on preserving chain-of-thought integrity when compressing distilled models to sub-4-bit and 8-bit precision.
• 5 min read
Releasing Project Solace 1.0 Omni: 12.59M Conversations Across 60 Datasets
We are releasing Project Solace 1.0 Omni, a 137.7 GB open dataset of verified frontier-model distillation conversations across 7 model families.