Direktori Model AI
Cari dan filter setiap model AI utama berdasarkan penyedia, harga API, jendela konteks, kecepatan, VRAM, dan benchmark.
Model AI Berdasarkan Perusahaan & Penyedia
0 match your filters
No models match your search. Try adjusting the filters.
Direktori Model AI — A–Z Index (330)
Direktori Model AI — FAQ
Pertanyaan Umum (FAQ)
How do you calculate the break-even token volume between self-hosted Ollama and managed APIs?+
Break-even occurs when Monthly GPU Infrastructure Cost (rental/depreciation + power + ops) equals Total Monthly Tokens multiplied by your Blended API Token Price; below 50M tokens/month, managed APIs are almost always cheaper.
Is it cheaper to rent an H100 GPU cluster or use DeepSeek V3 managed API?+
At $0.14 per 1M input tokens, DeepSeek V3 managed API is vastly cheaper than renting dedicated 8x H100 GPU clusters ($15-$25/hour) unless your organization sustains over 500 million tokens per day with 70%+ GPU utilization.
Why does low GPU utilization make self-hosting LLMs more expensive than API pay-as-you-go?+
Dedicated GPUs incur fixed costs 24/7; at 10% utilization during idle hours, your effective cost per processed token becomes 10 times higher than nominal capacity, erasing any theoretical hardware cost advantage.
What is the difference between cache write premium and cache read discount in Anthropic API?+
Anthropic charges a 1.25x (5-min TTL) or 2x (1-hour TTL) write premium on the initial request to compute the KV cache, then provides a 90% discount ($0.30/1M tokens) on all subsequent cache hit reads.
How do reasoning tokens affect total API invoice costs in OpenAI o3-mini?+
Reasoning tokens in o3-mini are billed at the standard output rate ($4.40/1M tokens); because complex reasoning steps can generate thousands of hidden scratchpad tokens, configuring reasoning_effort: low or medium is vital to prevent unexpected invoice inflation.
What is the difference between SWE-bench Verified and SWE-bench Pro leaderboards?+
SWE-bench Verified features 500 human-filtered GitHub issues where top models have reached 70-85%+ saturation, whereas SWE-bench Pro introduces harder multi-repo architectural tasks with strict contamination-proofing to evaluate autonomous engineering.
Can DeepSeek R1 run locally on a 24GB VRAM GPU like RTX 4090?+
The full 671B DeepSeek R1 requires multi-GPU clusters, but distilled versions (such as DeepSeek-R1-Distill-Qwen-14B or 32B with 4-bit quantization) run smoothly on a single 24GB RTX 4090.
Which model has lower latency for real-time voice streaming: Gemini 2.0 Flash or GPT-4o-mini?+
Gemini 2.0 Flash leads with native multimodal audio-to-audio streaming and sub-200ms time-to-first-token (TTFT), making it faster than text-pipelined GPT-4o-mini setups.