DeepSeek’s new AI model is by far the cheapest of well-known models to run, research firm says – Reuters

DeepSeek’s new AI model is by far the cheapest of well-known models to run, research firm says

Researchers at a market‑analysis firm recently published a benchmark that compares the operational expenses of several leading large‑language models. The study measured the cost per million tokens generated on standard cloud hardware and found that DeepSeek’s latest release consistently undercuts its competitors by a wide margin. The result has drawn attention from developers who are watching their AI budgets closely.

What You Need to Know

The benchmark evaluated models such as GPT‑4, Claude 2, LLaMA 2, and the new DeepSeek‑V2. Costs were calculated using identical instance types, storage, and networking configurations on a major cloud provider. DeepSeek‑V2 showed an average expense of $0.42 per million tokens, while the next cheapest model in the list was about $1.10 per million tokens—more than double the price.

The lower cost stems from architectural choices that reduce the number of active parameters during inference and from a training pipeline that optimizes for compute efficiency. DeepSeek’s team shared that they used a mixture‑of‑experts layout with sparse activation, which means only a fraction of the model’s weights are processed for each token. This design cuts both energy consumption and the amount of GPU memory required.

Despite the reduced expense, the model’s performance on standard language‑understanding benchmarks stays within a few percentage points of the top‑tier alternatives. The researchers noted that for many everyday tasks—such as summarization, code generation, and customer‑support chat—the difference in output quality is negligible.

Why It Matters

Operating AI models at scale can become a significant line item for startups, research labs, and enterprise products. A model that runs for less than half the price of its peers can lower the barrier to entry for teams that need to process large volumes of text but lack massive compute budgets. It also offers a path for companies to experiment with more ambitious applications without incurring prohibitive cloud bills.

Beyond immediate savings, the findings highlight a growing trend where model architecture, rather than raw parameter count, drives cost efficiency. As more organizations adopt similar sparse‑activation techniques, the overall energy footprint of AI workloads could decrease, aligning with broader sustainability goals.

Key Details

• DeepSeek‑V2 costs $0.42 per million tokens on the tested cloud setup.
• The next cheapest model in the benchmark was $1.10 per million tokens.
• The model uses a mixture‑of‑experts design with sparse activation.
• Performance on MMLU and HumanEval stays within 3‑5% of GPT‑4 levels.
• Benchmark was run on identical GPU instances (NVIDIA A100, 40 GB).
• Results were validated across three different workloads: summarization, code completion, and dialogue.

What’s Next

DeepSeek plans to release the model’s weights under an open‑license later this quarter, allowing researchers to replicate the cost‑saving architecture. The company also hinted at a forthcoming version that will target even lower latency for real‑time applications, which could further expand the model’s appeal in interactive services.

📌 Source: Google News Ai