CN — TechNode
DeepSeek Wants Better AI at Lower Cost
AI1 min read
DeepSeek, a Chinese AI company, updated its V4.1 Flash model to do tasks faster and cheaper. It also shared open-source software that helps run AI models more efficiently.
A Better Model at Lower Cost
On October 17, DeepSeek released V4.1-Flash-1017, an update to its V4.1 Flash model. The company says it scores higher on coding and agent tasks while using fewer tokens and less computing time. On October 20, it also released DeepInfer, an internal tool for running AI models, as open-source code.
Key Numbers
- Terminal-Bench 4.0 score rose from 31.2 to 42.3 (company's own test).
- DeepSWE v1.1 score rose from 74.2 to 77.0.
- On DeepSWE v1.1, the model scores about 77% at roughly $0.20 per task.
- At lower-cost settings of $0.11 to $0.14 per task, it still scores about 72% to 73%.
- On Arena's Agent Arena leaderboard, it ranked first among open-weight models and 11th overall, based on 7,498 real-use sessions. Median cost was about $0.18 per task.
What Is an Inference Framework?
Every time an AI model answers a question, it is doing inference. An inference framework decides how the model uses computer chips to handle these requests. Good design can make answers faster, let more users connect at once, and need less hardware. That makes it a big, though less visible, part of the cost of running AI.
Why It Matters
Making AI smarter is only part of the job. Companies also need to make it cheap enough to serve millions of people. DeepSeek's moves show it is focusing on both the model and the tools behind it. However, most of the benchmark results come from the company itself, so real results may vary. The open-source code is also not a full system yet, so developers still need extra engineering work.
What to Watch
- Whether DeepSeek's test results hold up in independent tests across different tasks.
- Whether developers adopt DeepInfer and how well it runs on different hardware.
- Per-request billing for DeepSeek's Claude Code and Codex integrations, scheduled to start on November 1.
- Whether DeepSeek adds the missing parts of a full inference system.
Glossary
- DeepSeek
- A Chinese AI company that builds large language models and offers them through an API.
- inference
- The step where a trained AI model is used to answer questions or complete tasks.
- DeepInfer
- DeepSeek's internal inference framework, now released as open-source code.
- open-weight models
- AI models whose trained settings (weights) are shared publicly so others can download and run them.