Tags give the ability to mark specific points in history as being important
-
helixllm-0.2.0
08de5c50 · ·feat(cpu): CPU-only inference — DeepSeek V4 models, hardware profiler, CPU env vars 5-phase implementation from 35-page CPU inference research: 1. DeepSeek V4 Flash (284B MoE, 120GB) + Pro (1.6T MoE, 680GB) in catalog 2. CPU hardware profiler: NUMA, memory bandwidth, L3, SIMD detection 3. Env vars: HELIX_INFERENCE_MODE, HELIX_CPU_THREADS, HELIX_NUMA/MMAP/MLOCK 4. llama-server CPU args: --numa, --mlock, --mmap, --no-kv-offload 5. Auto model selection: filter by RAM, prefer MoE (lowest active:total ratio) Branch: feature/helixllm-full-extension (3 commits merged to main)
-
helix-code-1.0.0-dev-0.0.1
071c1223 · ·helix-code-1.0.0-dev-0.0.1 — HelixLLM full-extension programme release Local RTX 5090 (Blackwell sm_120, CUDA 12.8) serving through HelixLLM -> HelixAgent -> HelixCode. Capabilities: embeddings, local vision, translation (NLLB), Whisper STT, Tesseract OCR, RAG (TEI), ACP->A2A, HelixAgent LAN/VPN network-provider, VRAM broker, LLMsVerifier chain, broad provider coverage. Governance (owned scope): cascade 0 failures, anti-bluff anchor present in all 18 files across 13 repos, §11.4.113 ff-only merge-onto-latest, §11.4.151 prefix. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
-
helixcode-v1.1.0
e035599d · ·helixcode-v1.1.0 release (§11.4.151 cross-repo marker) — component of the HelixCode v1.1.0 release