Glossary

Edited by Lawrence Beckwith on August 24, 2026 at 1:01 AM UTC

Short definitions for the terms used across these docs. Each links to the page that covers it properly.

Term Meaning
Ember OtherU's own inference engine — a narrow C/C++ runtime for DeepSeek V4 Flash on AMD Strix Halo, not a general GGUF loader. See Ember.
ds4 The upstream reference architecture (antirez/ds4) that Ember reimplements. Ember is a separate codebase, not a fork.
Hermes Agent The agent layer: channel routing (Telegram, Signal, WhatsApp), tool dispatch, memory and scheduling. Ember provides inference; Hermes decides what to do. See Connect Hermes Agent.
Hindsight The persistent memory system from Vectorize that gives Hermes long-term recall. See Hindsight Memory.
Memory bank A Hindsight container for memories; the API is scoped per bank (/v1/default/banks/{bank_id}/...).
DSpark The small draft model used for speculative decoding — it proposes tokens that the main model verifies, raising decode throughput.
Speculative decoding Running a small drafter ahead of the main model and accepting its tokens when they match, so several tokens can be produced per full forward pass.
Acceptance rate The fraction of drafted tokens the main model accepts. Workload-dependent — high on predictable text, lower on free-form sampling.
Abliteration A targeted edit that removes a model's refusal direction while leaving general capability intact. See Quantization Pipeline.
Quantization Storing model weights at reduced precision to fit in memory. OtherU ships DeepSeek-V4-Flash at 2.58 bits per weight.
bpw Bits per weight — the average precision of the quantized model.
imatrix An importance matrix collected from sample prompts, used to decide which weights tolerate the most aggressive quantization.
affine fp2 OtherU's 2-bit format, Q2_0_ROCMFP2, where value = code · scale − offset.
ROCmFPx The family of custom low-precision tensor formats Ember understands. Mainline llama.cpp, LM Studio and Ollama do not implement them.
MoE Mixture of Experts — only a subset of "expert" weights runs per token, so a large model can serve at a smaller active cost.
KV cache Cached attention state for tokens already processed, so a continued conversation need not re-process the whole prompt. Ember can page it to disk.
Prefill Processing the prompt before any output token is produced. Measured in tokens per second.
Decode Generating output tokens one at a time after prefill. Usually the slower phase.
gfx1151 The GPU architecture identifier for the AMD Radeon 8060S iGPU in Strix Halo — what Ember's kernels target.
Strix Halo The AMD Ryzen AI MAX+ platform with a large unified-memory iGPU, used as the reference hardware. See Hardware.
JetKVM The hardware KVM device that lets the agent see and control a physical machine over WebRTC capture and USB HID, with no software installed on the target.
GGUF The single-file model format Ember loads.