DeepSeek 发布 V4.1-Flash-1017，中美开源权重模型竞争升温

가

The body content of the article will be changed to this letter size.

스크롤 이동 상태바

Chinese artificial intelligence (AI) startup DeepSeek has released DeepSeek-V4.1-Flash-1017, an updated version of its V4.1 Flash model with improved processing speed, approximately five weeks after the original launch.

According to the model card published on Hugging Face, the release is classified as a minor update focused on reducing task execution time and output length. DeepSeek initially emphasized aggressive pricing, setting application programming interface (API) input costs at 15 cents per million tokens when the model debuted in September. The updated version costs $0.30 per million input tokens and $1.20 per million output tokens.

In DeepSeek’s internal testing, the updated model scored 77% on the DeepSWE v1.1 software engineering benchmark, up from 74.2% for the September version and above Anthropic’s Claude Opus 5 score of 74%. However, Anthropic’s model still outperformed DeepSeek in seven of the 11 major benchmarks disclosed by the Chinese company.

Independent AI model evaluation platform Artificial Analysis (AA) assigned DeepSeek-V4.1-Flash-1017 an Intelligence Index score of 45, placing it among the top 10 models. The model also ranked first in processing speed, outperforming second-place Meta Muse Wonder, which scored 176, by 54 points.

1. Artificial Analysis’ Intelligence Index evaluation of DeepSeek-V4.1-Flash-1017. (Source: AA)

Artificial Analysis’ processing speed evaluation of DeepSeek-V4.1-Flash-1017. (Source: AA)

Its cost per task was $0.26, the second lowest among the top 10 models in AA’s Intelligence Index.

Xiaomi’s MiMo-V2.6-Pro currently holds a narrow advantage over DeepSeek-V4.1-Flash-1017 in both cost and intelligence score. However, it ranks last in processing speed among the top 10 models. MiMo-V2.6-Pro is an open-weight model released by Xiaomi in September.

AA described DeepSeek-V4.1-Flash-1017 as reasonably priced and extremely fast, although somewhat verbose. The model generated 210 million tokens during AA’s Intelligence Index evaluation tasks, substantially above the median of 140 million tokens.

DeepSeek said the update was optimized for major agent software environments, including its own DeepSeek Harness, Tencent’s CodeBuddy AI coding assistant and WorkBuddy desktop agent.

The company also introduced two additional APIs for developers: a Search API that returns up to 50 web results per query and a Fetch API that converts individual web pages into Markdown documents. Both services cost $0.75 per 1,000 requests.

Artificial Analysis’ cost-per-task evaluation of DeepSeek-V4.1-Flash-1017. (Source: AA)

Meanwhile, the rapid advancement of Chinese AI models following Moonshot AI’s Kimi K3 has prompted growing calls within the US AI industry and academic community to develop domestic open-weight models to maintain the country’s leadership in AI.

Several US-developed open-weight models have been announced in recent days, including Reflection AI’s Beam, Nvidia’s Nemotron 3.5 Super and Volt, and Meta’s Muse Spark Open.

Meta’s release of Muse Spark Open marks its return to the open-weight ecosystem after Llama. The company had announced when Muse Spark was first launched that it would release the model weights.

According to Meta, Muse Spark Open has narrowed the performance gap with leading Chinese models and surpassed some of them. Meta also claims it delivers the strongest performance among currently available US-developed open-weight models.

Muse Spark Open uses a mixture-of-experts (MoE) architecture with 816 billion total parameters, of which 42 billion are active. It supports a context window of one million tokens.

Optimization for Apple’s MLX framework has also substantially improved execution efficiency on local Apple devices, including Mac Studio.

Artificial Analysis’ Intelligence Index rankings. Gray represents proprietary models, while shades of blue indicate open-weight models. Dark blue identifies open-weight models with restrictions on commercial use. Red arrows indicate, from left, DeepSeek-V4.1-Flash-1017, Meta Muse Spark, Reflection AI Beam and Nvidia Nemotron 3.5 Super. (Source: AA)

Reflection AI’s Beam is designed specifically for agentic task execution and coding.

While the article notes that many open-weight models use MIT licenses with commercial-use restrictions, Beam is released under the Apache 2.0 open-source license, lowering barriers to customization and commercial service development.

The model is regarded as an alternative combining technical capabilities with developer autonomy, following Meta’s approach. Nvidia is among Reflection AI’s investors.

Nvidia, which has been actively promoting open-weight models, also unveiled its 120-billion-parameter Nemotron 3.5 Super and Nemotron 3.5 Volt for lightweight agents on Oct. 15 local time.

Nemotron 3.5 Super combines Mamba-2 and Transformer architectures in an MoE design, activating only 12 billion of its 120 billion parameters per token.

The model uses Nvidia’s NVFP4 four-bit quantization technology to minimize performance degradation compared with the original model while achieving inference speeds of approximately 630 tokens per second with a 100,000-token prompt. It supports a context window of up to one million tokens.

Nemotron 3.5 Volt, designed for lightweight AI models with fewer than 100 billion parameters, activates 9 billion of its 74 billion total parameters and supports a 256,000-token context window.

Through reduced Top-K expert selection and optimized compression techniques, the model can run locally on a single graphics processing unit (GPU) requiring 24GB of video memory.

Despite having approximately 39% fewer parameters than its predecessor, Nemotron 3 Super, the new model has significantly improved its scores on major benchmarks. It also recorded the highest performance among US-developed open-weight models with fewer than 100 billion parameters.

Hyun-Seon, Park hspark@thelec.kr

My comment