The KV-cache reductions are especially important for deployment economics. Smaller active-parameter counts and lower per-token cache footprints can translate into materially better GPU memory utilization, denser multi-tenant inference, and cheaper long-context serving. But those gains depend on the maturity of inference stacks for the new architecture. Open weights under MIT are attractive, yet enterprises considering self-hosting still need to validate scheduler support, quantization behavior, SSD cache performance, and observability around prompt-prefill versus generation phases.
There is also a governance angle that the benchmark table does not answer. Automatic model substitution is convenient for cost control, but regulated or quality-sensitive workloads should treat it as a change event. Teams should re-run regression suites for coding agents, multimodal prompts and scientific reasoning workflows, especially where benchmark wins are narrow or domain-specific.
A practical response is straightforward:
- separate model selection from application logic via an internal gateway;
- track latency, token cost and task success by workload, not headline benchmark;
- add canary tests whenever a provider silently upgrades or reroutes a model family.
That discipline matters more than the leaderboard, because the real enterprise question is whether cheaper inference remains predictable enough to trust in production.
Chinese artificial intelligence startup Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd. today released DeepSeek-V4.1-Flash, the smallest model in a new architecture family. The company said tests by multiple parties put the open-weight model ahead of its much larger DeepSeek-V4-Pro on performance, cost, speed and total runtime.
Starting Sept. 14, requests sent to V4-Pro through DeepSeek’s application programming interface will be answered by V4.1-Flash and billed at the smaller model’s rates until a V4.1-Pro version launches. V4-Flash and the experimental vision model DeepSeek shipped in August are both retired. Calls to either now land on V4.1-Flash.
V4.1-Flash is a mixture-of-experts model with 552 billion parameters, close to double the 284 billion in V4-Flash. A new causal encoder-decoder design keeps just 8 billion parameters active while the model processes a prompt and 16 billion while it generates output. Image understanding, offered only in that experimental release last month, is now built into the model itself.
Much of the engineering went into shrinking the key-value cache. According to DeepSeek’s technical report, the model stores those entries in a four-bit floating-point format, and its global footprint comes to 890 bytes per token, about a quarter of what V4-Flash needs. Persistent cache storage on SSDs drops to roughly an eighth of the previous generation’s.
DeepSeek’s own benchmark table compares the model at maximum reasoning effort against Anthropic PBC’s Claude Opus 5 and OpenAI Group PBC’s GPT-5.6 Sol. V4.1-Flash scored 90.6 on Terminal-Bench 2.1, narrowly ahead of Opus 5 at 89.1 and GPT-5.6 Sol at 88.8. On the DeepSWE v1.1 software engineering test it resolved 74.2% of tasks, compared with 74% for Opus 5 and 62.7% for V4-Pro. Both U.S. models still lead on the GPQA Diamond science reasoning benchmark.
Off-peak API pricing is 15 cents per million uncached input tokens and 60 cents per million output tokens, with rates doubling during weekday peak windows. Developers still calling V4-Pro pay $3.96 per million output tokens at peak, compared with $1.20 for V4.1-Flash, so the rerouting works out to a cut of roughly 70% on output.
Weights are available on Hugging Face under the MIT license, and the model is live in DeepSeek’s web and mobile apps. DeepSeek said it will work with the open-source community on inference support and explore further deployment options.
The launch comes the same day Anthropic named DeepSeek in its latest threat intelligence report as one of seven China-based labs it says ran distillation campaigns against Claude. Anthropic attributed more than 12.1 million exchanges over 14 days in July to DeepSeek.
DeepSeek grew out of Chinese hedge fund High-Flyer. Founder Liang Wenfeng reportedly contributed $3 billion to a funding round of more than $7.4 billion in June that valued the company above $50 billion.
DeepSeek releases V4.1-Flash, says it outperforms flagship V4-Pro
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.