server/model_ranking.py

"""Все модели NVIDIA с ранжированием"""

from .config import HF_FALLBACK_MODEL

MODEL_RANKING = {
    # TIER 1: ELITE CODING
    "deepseek-v4-pro": {
        "endpoint": "deepseek-ai/deepseek-v4-pro",
        "context_window": 64000, "max_tokens": 8000,
        "coding_rank": 1, "speed_rank": 15, "reasoning_rank": 1,
        "cost_per_1k_input": 0.001, "cost_per_1k_output": 0.005,
        "tags": ["elite", "coding", "reasoning", "math", "cheap"]
    },
    "kimi-k2.6": {
        "endpoint": "moonshotai/kimi-k2.6",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 2, "speed_rank": 12, "reasoning_rank": 2,
        "cost_per_1k_input": 0.003, "cost_per_1k_output": 0.015,
        "tags": ["elite", "coding", "reasoning", "long_context"]
    },
    "qwen3.5-397b": {
        "endpoint": "qwen/qwen3.5-397b-a17b",
        "context_window": 128000, "max_tokens": 8000,
        "coding_rank": 3, "speed_rank": 18, "reasoning_rank": 3,
        "cost_per_1k_input": 0.002, "cost_per_1k_output": 0.008,
        "tags": ["elite", "coding", "long_context", "chinese"]
    },
    "mistral-large-3": {
        "endpoint": "mistralai/mistral-large-3-675b-instruct-2512",
        "context_window": 128000, "max_tokens": 8000,
        "coding_rank": 4, "speed_rank": 14, "reasoning_rank": 4,
        "cost_per_1k_input": 0.002, "cost_per_1k_output": 0.010,
        "tags": ["elite", "coding", "multilingual", "long_context"]
    },
    "gpt-oss-120b": {
        "endpoint": "openai/gpt-oss-120b",
        "context_window": 128000, "max_tokens": 8000,
        "coding_rank": 5, "speed_rank": 20, "reasoning_rank": 5,
        "cost_per_1k_input": 0.003, "cost_per_1k_output": 0.012,
        "tags": ["elite", "coding", "reasoning", "openai"]
    },

    # TIER 2: STRONG CODING
    "deepseek-v4-flash": {
        "endpoint": "deepseek-ai/deepseek-v4-flash",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 6, "speed_rank": 8, "reasoning_rank": 8,
        "cost_per_1k_input": 0.0005, "cost_per_1k_output": 0.002,
        "tags": ["strong", "coding", "fast", "cheap"]
    },
    "llama-4-maverick": {
        "endpoint": "meta/llama-4-maverick-17b-128e-instruct",
        "context_window": 128000, "max_tokens": 8000,
        "coding_rank": 7, "speed_rank": 10, "reasoning_rank": 7,
        "cost_per_1k_input": 0.001, "cost_per_1k_output": 0.004,
        "tags": ["strong", "coding", "meta", "long_context"]
    },
    "nemotron-3-super": {
        "endpoint": "nvidia/nemotron-3-super-120b-a12b",
        "context_window": 128000, "max_tokens": 8000,
        "coding_rank": 8, "speed_rank": 16, "reasoning_rank": 6,
        "cost_per_1k_input": 0.002, "cost_per_1k_output": 0.008,
        "tags": ["strong", "coding", "nvidia", "reasoning"]
    },
    "mistral-medium-3.5": {
        "endpoint": "mistralai/mistral-medium-3.5-128b",
        "context_window": 64000, "max_tokens": 8000,
        "coding_rank": 9, "speed_rank": 11, "reasoning_rank": 10,
        "cost_per_1k_input": 0.001, "cost_per_1k_output": 0.005,
        "tags": ["strong", "coding", "mistral", "balanced"]
    },
    "dracarys-llama-70b": {
        "endpoint": "abacusai/dracarys-llama-3.1-70b-instruct",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 10, "speed_rank": 13, "reasoning_rank": 11,
        "cost_per_1k_input": 0.001, "cost_per_1k_output": 0.004,
        "tags": ["strong", "coding", "roleplay", "creative"]
    },
    "llama-3.3-70b": {
        "endpoint": "meta/llama-3.3-70b-instruct",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 11, "speed_rank": 9, "reasoning_rank": 12,
        "cost_per_1k_input": 0.0005, "cost_per_1k_output": 0.002,
        "tags": ["strong", "coding", "meta", "fast", "cheap"]
    },
    "nemotron-super-49b": {
        "endpoint": "nvidia/llama-3.3-nemotron-super-49b-v1.5",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 12, "speed_rank": 7, "reasoning_rank": 13,
        "cost_per_1k_input": 0.0005, "cost_per_1k_output": 0.002,
        "tags": ["strong", "coding", "nvidia", "fast", "cheap"]
    },

    # TIER 3: GOOD CODING
    "step-3.7-flash": {
        "endpoint": "stepfun-ai/step-3.7-flash",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 13, "speed_rank": 6, "reasoning_rank": 14,
        "cost_per_1k_input": 0.0005, "cost_per_1k_output": 0.002,
        "tags": ["good", "coding", "fast", "chinese", "cheap"]
    },
    "mistral-small-4": {
        "endpoint": "mistralai/mistral-small-4-119b-2603",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 14, "speed_rank": 5, "reasoning_rank": 15,
        "cost_per_1k_input": 0.0005, "cost_per_1k_output": 0.002,
        "tags": ["good", "coding", "fast", "mistral", "cheap"]
    },
    "minimax-m2.7": {
        "endpoint": "minimaxai/minimax-m2.7",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 15, "speed_rank": 4, "reasoning_rank": 16,
        "cost_per_1k_input": 0.0005, "cost_per_1k_output": 0.002,
        "tags": ["good", "coding", "fast", "chinese", "cheap"]
    },
    "nemotron-super-49b-v1": {
        "endpoint": "nvidia/llama-3.3-nemotron-super-49b-v1",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 16, "speed_rank": 17, "reasoning_rank": 17,
        "cost_per_1k_input": 0.0005, "cost_per_1k_output": 0.002,
        "tags": ["good", "coding", "nvidia", "cheap"]
    },
    "llama-3.2-90b-vision": {
        "endpoint": "meta/llama-3.2-90b-vision-instruct",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 17, "speed_rank": 19, "reasoning_rank": 18,
        "cost_per_1k_input": 0.001, "cost_per_1k_output": 0.004,
        "tags": ["good", "coding", "vision", "multimodal", "meta"]
    },

    # TIER 4: FAST / LIGHT
    "nemotron-nano-12b": {
        "endpoint": "nvidia/nemotron-nano-12b-v2-vl",
        "context_window": 16000, "max_tokens": 4000,
        "coding_rank": 18, "speed_rank": 2, "reasoning_rank": 22,
        "cost_per_1k_input": 0.0001, "cost_per_1k_output": 0.0005,
        "tags": ["light", "fast", "vision", "nvidia", "cheap"]
    },
    "nemotron-3-nano-30b": {
        "endpoint": "nvidia/nemotron-3-nano-30b-a3b",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 19, "speed_rank": 3, "reasoning_rank": 19,
        "cost_per_1k_input": 0.0002, "cost_per_1k_output": 0.001,
        "tags": ["light", "fast", "nvidia", "cheap"]
    },
    "nemotron-nano-9b": {
        "endpoint": "nvidia/nvidia-nemotron-nano-9b-v2",
        "context_window": 16000, "max_tokens": 4000,
        "coding_rank": 20, "speed_rank": 1, "reasoning_rank": 23,
        "cost_per_1k_input": 0.0001, "cost_per_1k_output": 0.0005,
        "tags": ["light", "fastest", "nvidia", "cheap"]
    },
    "nemotron-content-safety": {
        "endpoint": "nvidia/nemotron-content-safety-reasoning-4b",
        "context_window": 8000, "max_tokens": 2000,
        "coding_rank": 21, "speed_rank": 1, "reasoning_rank": 24,
        "cost_per_1k_input": 0.0001, "cost_per_1k_output": 0.0005,
        "tags": ["light", "fastest", "safety", "nvidia", "cheap"]
    },

    # TIER 5: SPECIALIZED
    "nemotron-3-nano-omni": {
        "endpoint": "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 22, "speed_rank": 5, "reasoning_rank": 20,
        "cost_per_1k_input": 0.0002, "cost_per_1k_output": 0.001,
        "tags": ["specialized", "omni", "multimodal", "reasoning", "nvidia", "cheap"]
    },
    "diffusiongemma": {
        "endpoint": "google/diffusiongemma-26b-a4b-it",
        "context_window": 16000, "max_tokens": 4000,
        "coding_rank": 23, "speed_rank": 10, "reasoning_rank": 25,
        "cost_per_1k_input": 0.001, "cost_per_1k_output": 0.004,
        "tags": ["specialized", "image", "diffusion", "google"]
    },

    # LEGACY
    "glm": {
        "endpoint": "z-ai/glm-5.1",
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 9, "speed_rank": 8, "reasoning_rank": 9,
        "cost_per_1k_input": 0.002, "cost_per_1k_output": 0.008,
        "tags": ["legacy", "coding", "fast", "chinese"]
    },
    "hf_fallback": {
        "endpoint": HF_FALLBACK_MODEL,
        "context_window": 32000, "max_tokens": 8000,
        "coding_rank": 50, "speed_rank": 50, "reasoning_rank": 50,
        "cost_per_1k_input": 0, "cost_per_1k_output": 0,
        "tags": ["fallback", "free", "hf"]
    },
}