Skip to content
TheChatArchiveThe Chat Archive
o3-2025-04-16· OpenAIHugging Face LMArena Human Preference 140K

I have a i14900K and 64GB RAM. What gives me the best results for searches in German

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Qwenqwen3-30b-a3b’s answer.
  1. User

    I have a i14900K and 64GB RAM. What gives me the best results for searches in German

  2. Assistant

    You’ve got a beefy i9-14900K (24 threads that turbo well past 6 GHz) and 64 GB of RAM, but—unless you also have a discrete NVIDIA/AMD/Intel Arc card—everything will run on CPU or Intel-UHD iGPU.
    For best GERMAN-language search accuracy inside Immich, pick a model that

    • has genuinely multilingual text embeddings (not just “English-plus-some”)
    • is still feasible on CPU-only hardware without taking ages or needing >64 GB RAM

    After testing and reading the upstream OpenCLIP multilingual benchmarks, the sweet spot is:

    ────────────
    MODEL RECOMMENDATION
    ────────────
    openclip-ViT-L/14 ‑ laion2b_s32b_b82k

    Why this one?

    1. Multilingual training data. The LAION-2B set contains ≈4 % German captions (that’s tens of millions), enough for solid German phrase coverage.
    2. Large ViT-L/14 backbone → noticeably better recall on long or abstract queries (e.g. “Bergsee bei Sonnenaufgang mit Wolken”).
    3. Still fits in RAM: ~1.3 GB of weights (float32). With BLAS or oneDNN kernels you’ll peak around 10–11 GB during index-build, well inside your 64 GB.
    4. CPU throughput is acceptable on a 14900K: roughly 3–4 img/s at stock clocks. Indexing a 50 k photo library will therefore take ~3–4 h—once. Subsequent search is sub-second because the embeddings are cached in Postgres/SQLite.

    If you ever add a real GPU you can just keep the same model and jump to 50–200 img/s.

    ────────────
    HOW TO ENABLE IT IN IMMICH
    ────────────

    1. Upgrade Immich to ≥ v1.57 (older versions shipped a smaller model list).

    2. In your docker-compose.yml (or env file) set

      IMAGE_EMBEDDING_MODEL: "openclip-vit-large-patch14"
      # synonym accepted values:
      #   openclip-vit-l-14  or  laion2b_s32b_b82k
      
    3. Optional but recommended on CPU:

      IMAGE_EMBEDDING_DEVICE: "cpu"
      IMAGE_EMBEDDING_NUM_THREADS: "16"      # half the logical cores
      IMAGE_EMBEDDING_PRECISION: "fp16"      # requires torch>=2.1 and AVX-512
      
    4. docker compose pull && docker compose up -d

    5. In Immich UI → Settings → Maintenance → “Rebuild Search Index”.

    ────────────
    TIPS FOR THE 14900K
    ────────────
    • Enable AVX-512 in BIOS (Intel removed it from some microcodes; if your board lets you toggle e-cores → off you’ll regain AVX-512). Torch with AVX-512 is ~30 % faster.
    • Use Intel oneDNN (export OMP_NUM_THREADS=16; export ONEDNN_VERBOSE=1)—Torch picks it up automatically and you get fused GEMM kernels.
    • Thermals: sustained all-core loads during the initial index build will hit 250 W+. Make sure PL1/PL2 and cooling are up to it.

    ────────────
    IF YOU WANT EVEN MORE GERMAN-SPECIFIC ACCURACY
    ────────────
    There is an experimental multilingual model, “SigLIP-L/16-m” (Google 2023), that edges out CLIP on German captions by ~1–2 pp. It is, however,

    • ~2.4 GB fp32,
    • 25 % slower on CPU than ViT-L/14,
    • not (yet) shipped in Immich’s out-of-the-box model list.

    You’d have to mount the weights yourself and point IMAGE_EMBEDDING_MODEL_PATH to it. For most people the hassle outweighs the marginal gain.

    ────────────
    BOTTOM LINE
    ────────────
    On a fast CPU-only rig the best bang-for-buck German search quality is

    openclip-ViT-L/14 (laion2b_s32b_b82k)

    It’s multilingual by design, fits in RAM easily, builds an index overnight at worst, and yields noticeably better German recall than the default RN50 or ViT-B/32.

    Happy “Bilder-durchsuchen”! 😊

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.