{
  "data": null,
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/niches/decision-models.run-on-own-hardware/",
    "json": "https://www.anchorterminal.com/niches/decision-models.run-on-own-hardware/index.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/niches/decision-models.run-on-own-hardware/index.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/niches/decision-models.run-on-own-hardware/index.min.md"
  },
  "markdown": "Two or three companies' docs show a full fit.\n\n| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |\n| --- | --- | --- | --- | --- |\n| 2 | 10 | 121 | 10 | 9 |\n\n## What counts as done\n\nThe documented server answers a Jev-shaped request on the stated hardware with the trained probabilities (not a causal-attention approximation), with authentication available.\n\nMeasured, for our own tests: under 1 s per decision.\n\nHow agents ask for it: run this classifier on my Mac; self-host a router so data does not leave; serve it on a T4 or CPU box; load it with vLLM or MLX\n\n## Does the job\n\nThe docs show it meets the test above, and nothing we read says otherwise.\n\n- **Laya (Convai Innovations Pvt. Ltd.)**: pip install laya, then laya-serve exposes /v1/systemone on CUDA, MPS or CPU with LAYA_API_KEY for a bearer key (it binds 0.0.0.0 by default); 33 ms on a T4, 193 to 464 ms on CPU, stock ModernBERT weights on Transformers.\n- **Vela 2.0 (vLLM Semantic Router project (with KR Labs))**: AutoModel with trust_remote_code on Transformers 5.17 or later, FP32 on CPU supported, vela2_serve.py answers /v1/systemone with VELA2_API_KEY for a bearer key and 413 when questions exceed the 16,384-token budget. Engine mode needs the router's dev channel.\n\n## Does the job, with a limit\n\nIt does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.\n\n- **Clef (Cloudflare, Inc.)**: Apache-2.0 weights with encode_record and a Jev / SystemOne section; tested with Transformers 5.10.2 on one H200 and served by SGLang. Self-hosting runs the repository's own code, and no bundled server with a key option is documented.\n- **Decider (Mark Marosi (Mapika))**: pip install decider-ai runs on CUDA, MPS, CPU (float32) and GGUF through llama-cpp-python, serving /v1/systemone on 127.0.0.1. The server has no API key option and cuts states over 32,768 tokens silently; issue 18 records Windows on ARM install and CPU speed problems.\n- **Decision 3.0 (vLLM Semantic Router project)**: vllm serve loads the repository natively and answers /v1/systemone, verified on ROCm, but on vLLM's nightly wheels; d3_server.py binds 127.0.0.1 with DECISION_API_KEY and refuses over-long input. Transformers needs trust_remote_code and Engine mode the router's dev channel.\n- **Drex Max v1 (Nace AI, Inc.)**: Every path that runs it as trained is Nace's code or fork: drex_max on CUDA, Kapitsa and MLX scripts on a 24 GB Mac, a drex-max-v1 branch of llama.cpp and Ollama. Stock loaders run the full-attention layers causally. The Kapitsa server's key option wasn't read.\n- **elephant-1 (Baseten Labs, Inc.)**: setup.sh clones OneJev at a pinned commit, applies serve.patch and installs the Decision Index kit on one CUDA GPU (about 19 GB of BF16 files). Stock vLLM and Transformers ignore the block-bidirectional mask. No bundled HTTP server or key option.\n- **Kev (Jared Palmer)**: Runs from a Git clone with uv (the PyPI kev is unrelated) on CUDA, ROCm, MLX and CPU; KEV_API_KEY requires a bearer key and states over 65,536 tokens get a 422. Kev-4B takes 721 ms for five questions on an Apple M5. Issue 281's VRAM growth is open.\n- **joey-1 (Baseten Labs, Inc.)**: setup.sh clones OneJev at a pinned commit, applies serve.patch and installs the Decision Index kit on one CUDA GPU (5.5 GB of BF16 files, an 8 GB GPU). Stock vLLM and Transformers ignore the block-bidirectional mask. No bundled HTTP server or key option.\n- **Liquid d1 (Liquid AI, Inc.)**: d1-2B serves /v1/systemone from llama.cpp 0.8.0 (llama serve -hf LiquidAI/d1-2B-GGUF), MLX builds and Transformers 5.19 with text, pictures and video. The page shows no key option, and the LFM licence ends free commercial use at $10 million revenue.\n- **mammoth-1 (Baseten)**: setup.sh clones OneJev at a pinned commit, applies serve.patch and installs the Decision Index kit on one CUDA GPU (about 55 GB of BF16 files, an 80 GB GPU). Stock vLLM and Transformers ignore the block-bidirectional mask. No bundled HTTP server or key option.\n- **pplx-decider (Perplexity AI, Inc.)**: The weights run through the repository's pplx_decider package on CUDA (about 52 GiB in BF16); the card says default causal inference does not reproduce the checkpoint. No bundled HTTP server or key option documented for the 27B.\n- **pplx-decider-flash (Perplexity AI, Inc.)**: Decider for Mac (Apple silicon, 16 GB, 5.8 GB download) serves the same JSON at 127.0.0.1:8080 with no API key, and the weights run on CUDA through pplx_decider (about 9 GiB). No authentication option on the Mac server.\n- **Strands Decider 2B (Amazon Web Services, Inc.)**: pip install strands-decider serves /v1/systemone on CUDA, MPS, CPU or MLX, on 127.0.0.1 with no authentication option; a prompt over the 4,096 window is refused with 422 under --strict-window and cut otherwise. 11 to 13 s a decision on a 2 vCPU box.\n\n## Couldn't check\n\nThe docs were closed to us, or silent. These carry no mark either way.\n\n- **Drex 1.5 (Nace AI, Inc.)**: The drex-v1.5 weights on Hugging Face weren't read in this pass, so the runtime path for this checkpoint is unverified; see the drex-max-v1 listing for the family's loader situation.\n- **kettei (Sqwish Ltd, trading as Sqwish Labs)**: The kettei core and dot weights under Sqwish's custom licence weren't read; the docs cover the hosted API only.\n\n- not this job: Celeris-1 Decision (Marqo Inc): Closed weights behind a hosted API. The docs offer no download or self-hosting path.\n- not this job: GLiClass (Knowledgator Engineering Ltd.): Runs locally on CPU or GPU and serves over Ray, but the server is not Jev-shaped (issue 47) and checks no credential, so the documented server can't answer a Jev-shaped request with authentication.\n- not this job: Granite Switch 4.1 (International Business Machines Corporation): Serves on released vLLM with the granite-switch package on the owner's GPU, but there is no Jev-shaped route, so the success criterion isn't met; the card says to add authentication before exposing vLLM.\n- not this job: JEV-27B-VL (AutoTrust AI Lab): serve.sh runs serve_decide.py on a September 2026 vLLM development build; released vLLM needs client-side arithmetic from decision_head.json. The /v1/decide shape differs from /v1/systemone, so a Jev-shaped request isn't answered.\n- not this job: Microsoft-Decision-1 (Microsoft): Hosted on Foundry and OpenRouter only; no open weights were found.\n- not this job: OpenAI Decisions API (OpenAI OpCo, LLC): Closed model behind the OpenAI API.\n- not this job: RightNow System Zero (RightNow AI, Inc.): Closed weights, hosted only.\n- not this job: ScaleDown (Carbon ScaleDown, Inc.): Hosted API only; no weights.\n- not this job: SPEV (Sprag AI, Inc.): Apache-2.0 30B-A3B weights load in Transformers with trust_remote_code, but no server is documented and the hosted shape is OpenAI's, so no documented server answers a Jev-shaped request.\n- not this job: Jev (TypeSafe AI, Inc.): Closed model behind the hosted API; the models page says the same weights serve every account.\n\n1 listings in the category were added after this reading and aren't marked yet.\n\n## What stops agents finishing\n\n- weights run as trained only on a patched or development runtime (mammoth-1, joey-1, Drex Max, JEV-27B-VL)\n- server is CUDA-only or segfaults on ROCm and Windows on ARM\n- 11 to 13 s per decision on a 2 vCPU box\n- bundled server has no API key option\n\n## Around it\n\nBefore: model-hubs.download-weights. After: decision-models.route-to-queue, decision-models.gate-action.\n\n## Where this comes from\n\nWeight 10 of 100 in Decision models: 9 of 41 accounts, mostly GitHub issues on CPU, MPS, ROCm, Windows on ARM, VRAM growth and patched runtimes\n\nAs data: https://www.anchorterminal.com/api/v1/use-cases/decision-models.run-on-own-hardware.json, or the MCP tool get_use_case with id decision-models.run-on-own-hardware. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-11-16",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.8",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "For agents",
        "url": "https://www.anchorterminal.com/agents/"
      },
      {
        "name": "Niches",
        "url": "https://www.anchorterminal.com/niches/"
      },
      {
        "name": "Run the decision model on the owner's own machine",
        "url": ""
      }
    ],
    "description": "Run the decision model on the owner's own machine: 2 companies fully do it, 10 with a limit, 121 carry the tag. Every listing in Decision models with the reason and a source.",
    "facts": null,
    "h1": "Run the decision model on the owner's own machine",
    "image": "https://www.anchorterminal.com/assets/og/default.png",
    "path": "/niches/decision-models.run-on-own-hardware/",
    "published": "",
    "section": "agents",
    "title": "Run the decision model on the owner's own machine | Anchor Terminal",
    "toc": null,
    "updated": "2026-11-15",
    "url": "https://www.anchorterminal.com/niches/decision-models.run-on-own-hardware/"
  },
  "tokens": {
    "markdown": 2200,
    "slim": 380
  },
  "version": 1
}
