Summary
On an Intel N95 (Alder Lake-N, E-cores only) running Windows, the prebuilt node-llama-cpp binaries shipped with OpenClaw's qmd integration crash with a Windows access violation (0xC0000005) the first time they try to load any GGUF model for embeddings/reranking/query-expansion. BM25-only paths work fine; only the LLM half is affected.
Environment
- OpenClaw: 2026.5.6
- Host: Intel N95 (4× E-cores, no P-cores), Windows 10/11 x64 (26200)
- Node: v24.15.0
node-llama-cpp: version installed by OpenClaw qmd integration as of 2026-05-07
- Model attempted: small GGUF embeddings + reranker shipped with qmd default config
Repro
- Install qmd / LLM half on a host with an Intel Alder Lake-N CPU (N95, N100, N200, N300).
- Index a workspace (BM25 builds successfully — 476 docs / ~1 s queries in our case).
- Trigger a query that exercises the LLM half (embedding lookup, reranker, or query expansion).
- Process exits with
0xC0000005 (Windows access violation) on first model load.
Expected
Model loads; embeddings/rerank/query-expansion work; hybrid search returns results.
Actual
Hard crash on model load. No useful stderr beyond the 0xC0000005 code. BM25-only path remains usable as a workaround.
Suspected cause
Prebuilt node-llama-cpp Windows binaries appear to assume P-core CPU features (likely AVX2 / specific SIMD ISA extensions) not present on Alder Lake-N E-cores. Source-built binaries via Visual Studio 2022 Build Tools are reported to work but require ~6 GB and ~45 min of install time, which is a steep ask for a small-form-factor user.
Impact
Memory recall on this hardware is degraded to BM25-only (works for direct keyword hits, fails on paraphrased queries). Several Mini-PC and N100/N95 form factors are popular for self-hosted home servers, so this is likely to affect a non-trivial slice of OpenClaw users on Windows.
Suggested fixes
- Detect Alder Lake-N at install/load and either fall back gracefully or build from source automatically.
- Ship a separate set of E-core-safe prebuilt binaries.
- Document the limitation prominently so users know to install Build Tools up front.
Summary
On an Intel N95 (Alder Lake-N, E-cores only) running Windows, the prebuilt
node-llama-cppbinaries shipped with OpenClaw's qmd integration crash with a Windows access violation (0xC0000005) the first time they try to load any GGUF model for embeddings/reranking/query-expansion. BM25-only paths work fine; only the LLM half is affected.Environment
node-llama-cpp: version installed by OpenClaw qmd integration as of 2026-05-07Repro
0xC0000005(Windows access violation) on first model load.Expected
Model loads; embeddings/rerank/query-expansion work; hybrid search returns results.
Actual
Hard crash on model load. No useful stderr beyond the
0xC0000005code. BM25-only path remains usable as a workaround.Suspected cause
Prebuilt
node-llama-cppWindows binaries appear to assume P-core CPU features (likely AVX2 / specific SIMD ISA extensions) not present on Alder Lake-N E-cores. Source-built binaries via Visual Studio 2022 Build Tools are reported to work but require ~6 GB and ~45 min of install time, which is a steep ask for a small-form-factor user.Impact
Memory recall on this hardware is degraded to BM25-only (works for direct keyword hits, fails on paraphrased queries). Several Mini-PC and N100/N95 form factors are popular for self-hosted home servers, so this is likely to affect a non-trivial slice of OpenClaw users on Windows.
Suggested fixes