DeepSeek V4 Pro Wins Second Place in Programming Evaluation: Score Only Slightly Lower than Kimi K3, but Cost is Only One Fourteenth of the Opponent

CLUE team's latest released SuperCLUE-Terminal Chinese AI terminal programming evaluation list shows that DeepSeek-V4-Pro-0813 scored 51.52, ranking second among all tested models, only behind the top-ranked Kimi K3. More notably, it maintains an extremely low call cost while achieving this score, making its cost-effectiveness particularly prominent among high-scoring models. This list is specifically designed for domestic developers, using all local real programming tasks and a unified running framework of Claude Code, fairly comparing major models' complete abilities in understanding Chinese requirements, planning tasks, calling tools, and debugging code.

The realism of the evaluation comes from its task settings: each question requires the model to complete 50 to 110 rounds of interaction, with a full run time of half an hour to one and a half hours, which highly replicates complex workflows in daily development. In this round of the list, Kimi K3 ranked first with 60.61 points, followed closely by DeepSeek-V4-Pro-0813, scoring higher than GLM-5.2's 48.48 and the old version of DeepSeek-V4-Flash's 46.46, firmly placing it in the top tier.

1.42 yuan per question cost, competing with top-tier computing power

The most impressive is DeepSeek-V4-Pro-0813's cost control: the call cost per question is only about 1.42 yuan, roughly one fourteenth of Kimi K3's and one sixteenth of GLM-5.2's. At a time when the scores of top models often differ by just a few points, this order-of-magnitude cost gap means that small and medium-sized teams can afford "nearly top-tier" coding capabilities without paying exorbitant computing costs for every interaction.

From actual testing scenarios, the model covers real demands such as front-end pages, 3D games, and engineering script modifications, able to fully close the game development logic, with collision, scoring, and settlement functions operating normally. The page color and interaction feedback also have moved away from the template-like AI style. There may be minor structural issues in some creative drawing questions, but the overall completion is still leading. For budget-sensitive small and medium enterprises and independent developers, DeepSeek-V4-Pro-0813 redefines the cost-effectiveness benchmark of programming models with a "second-place score and one-fourteenth price"—when top-tier capabilities no longer mean top-tier expenses, the entry barrier for AI programming is truly being lowered.

Nighttime Shock Release: DeepSeek V4 Pro Official Version Makes Its Debut, Standing Shoulder to Shoulder with Top Fable Models

On August 13th, there was a concentrated release of AI products, with Musk launching Grok4.6 and Qwen3.8Max being open-sourced. DeepSeek simultaneously quietly launched the official version of V4 Pro, DeepSeek-V4-Pro-0813. The API and Chat endpoints have been updated, and users can access it by specifying the name as before. Official benchmark scores show its core indicators perform excellently.

From Trillion-Code Open Source to Domestic Computing Power Collaboration: Kimi K3 and MoLi Thread Intelligent Computing Card Complete Full Stack Adaptation

Moonshot AI open-sourced the 2.8T-parameter Kimi K3 model on July 28, supporting 1M-token context and vision, with KDA hybrid attention. Moore Threads achieved fast adaptation and stable operation on its MTT S5000 smart computing card, marking a breakthrough in domestic computing power and open-source LLM synergy.....

China's GPU Landscape Gains a New Powerhouse: Moerlian Successfully Runs Open-Source Giant Model with 2.8 Trillion Parameters on Kimi K3 MUSA Stack

Moore Threads announced it has adapted the open-sourced 2.8-trillion-parameter Kimi K3 model from Moonshot AI, achieving fast and stable operation on its self-developed MTT S5000 training-inference card and MUSA software stack, bolstering domestic computing support for trillion-parameter models.....

Domestic GPUs Can Also Run Trillion-Parameter Large Models: Hygon DCU Adapts to Kimi K3 896 Expert Parallelism Without Lag

Hygon DCU has completed full-stack adaptation and verification of the trillion-parameter large model Kimi K3, marking the first time domestic computing power has stably supported such a massive model. Through deep optimization from the underlying operators to the inference engine, the model requires no code changes, making the migration cost almost zero. Developers can deploy the model as soon as they obtain computing power. Meanwhile, it achieves seamless compatibility with KDA attention and the 896 expert MoE architecture, breaking the monopoly of overseas flagship chips and providing an out-of-the-box domestic solution.

At least 18 US companies quickly deployed Kimi K3 on its open-source night

After Kimi K3 weights were open-sourced on Hugging Face, at least 18 US companies swiftly deployed them commercially, contradicting US government lobbying against Chinese open-source models. The list, compiled by user Ding, exposes a rift between Washington's political narrative and Silicon Valley's pragmatic rationality, where top performance and cost efficiency override political posturing.....