API and command line

Command line

With uv installed, run it without installing anything:

Source code: github.com/theaibench/theaibench-cli (MIT). Also on PyPI as theaibench.

HTTP API

Returns JSON. Browsers that open the link are sent to the checker with the same settings. Add format=json to force JSON. CORS is open, so web apps can call it directly.

Parameters

Allowed vram: none, 8, 12, 16, 20, 24, 32, 48, 64, 96, 128. Allowed context: 4096, 8192, 16384, 32768, 65536, 131072, 200000.

Example response

This is the real answer for the request above, made by the same code the site uses:

{
  "inputs": {
    "mode": "current",
    "platform": "windows",
    "pcVram": "24",
    "macMemory": "32",
    "systemRam": "32",
    "budget": "1500to3000",
    "useCase": "coding",
    "priority": "speed",
    "gpuFamily": "nvidia",
    "context": "32768",
    "chip": null
  },
  "result": {
    "verdict": "Comfortable",
    "tier": 5.27,
    "band": "high",
    "title": "Best fit: Qwen3.8 27B",
    "summary": "24 GB NVIDIA graphics card (RTX 4090-class speed) runs 20 of 26 model families we track for coding. Best: Qwen3.8 27B (Q4_K_M).",
    "picks": [
      {
        "name": "Qwen3.8 27B",
        "why": "Q4_K_M, needs 21.5 GB at 32K context · ~30–50 tok/s (estimate) · ollama run qwen3.8:27b",
        "slug": "qwen-3-8",
        "url": "https://theaibench.ai/models/qwen-3-8/",
        "fit": "good",
        "quant": "Q4_K_M",
        "memory_gb": 21.5,
        "speed_tok_s": {
          "low": 30,
          "high": 48
        },
        "ollama": "qwen3.8:27b"
      },
      {
        "name": "Muse Glimmer 30B",
        "why": "Q4_K_M, needs 20.2 GB at 32K context · ~35–50 tok/s (estimate) · ollama run muse-glimmer:30b",
        "slug": "muse-glimmer",
        "url": "https://theaibench.ai/models/muse-glimmer/",
        "fit": "good",
        "quant": "Q4_K_M",
        "memory_gb": 20.2,
        "speed_tok_s": {
          "low": 33,
          "high": 52
        },
        "ollama": "muse-glimmer:30b"
      },
      {
        "name": "Qwen3.6 35B-A3B",
        "why": "UD-IQ4_XS, needs 20.3 GB at 32K context · ~110–210 tok/s (estimate)",
        "slug": "qwen-3-6",
        "url": "https://theaibench.ai/models/qwen-3-6/",
        "fit": "good",
        "quant": "UD-IQ4_XS",
        "memory_gb": 20.3,
        "speed_tok_s": {
          "low": 106,
          "high": 209
        },
        "ollama": null
      }
    ],
    "runner": {
      "name": "Ollama",
      "note": "ollama run qwen3.8:27b"
    },
    "quantization": "Q4_K_M for most models; Q8_0 when memory allows. gpt-oss ships as MXFP4.",
    "expected_speed": "~30–50 tok/s for Qwen3.8 27B (estimate from memory bandwidth)",
    "workflow": [
      "Install Ollama (ollama.com/download) or LM Studio.",
      "Run: ollama run qwen3.8:27b",
      "Apps can use it at http://localhost:11434 (Ollama) as an OpenAI-style API."
    ],
    "watchouts": [],
    "note": "Try other settings at https://theaibench.ai/?hw=rtx-4090&mem=24&ram=32&task=coding&ctx=32768",
    "checker_url": "https://theaibench.ai/?hw=rtx-4090&mem=24&ram=32&task=coding&ctx=32768"
  },
  "meta": {
    "version": "v1",
    "engine": "v2",
    "dated": "October 2026",
    "source": "https://theaibench.ai/",
    "docs": "https://theaibench.ai/api/",
    "method": "https://theaibench.ai/about/#method",
    "license": "Free to cite with attribution to The AI Bench."
  }
}

Chip ids

Other machine-readable files

• OpenAPI description — for GPT Actions and tool-calling agents.
• llms.txt — a plain-text map of the site for AI assistants.

Fair use

Please cache answers and keep requests reasonable. Credit “The AI Bench” with a link when you show the results.