DeepSeek C.E Flash 1017
An open-source large model with a mixture-of-experts architecture released by DeepSeek, with 552 billion total parameters and 16 billion activated parameters, optimized specifically for agent tasks.
Timeline (last 90 days)
The model has been open-sourced on Hugging Face and ModelScope, totaling about 510GB, and the community has already released quantized versions (e.g., Unsloth's 4bit GGUF at only 233GB).
On benchmarks for agentic coding, general agents, knowledge, automation and others, DeepSeek C.E Flash 1017 improves by several percentage points over the previous generation, beats open-source models such as GLM5.3 and Kimi K3, and performs almost on par with GPT-5.6 Sol.
On KernelBench (low-level GPU code writing), DeepSeek C.E Flash 1017 ranks third, behind only the latest Claude model, but beats Gemini 4 Argon, Claude Fable and GPT Leo Astra.
In the UP主's hands-on test, DeepSeek C.E Flash 1017 ran for 36 minutes to complete the task of explaining the measurement of Earth's circumference, consuming about 22 million tokens with a 99% cache hit rate.
In the Unreal Engine procedural 3D game task, the model ran 16 rounds, and the critic agent gave it a maximum score of 5.5 out of 10, falling short of the 8.5 AAA standard, consuming about 375 million tokens with a 99.5% cache hit rate.
In the music creation test, the model's first-round result was poor, and the second-round melody was similar to other models on the market; the UP主 considered it inferior to GPT-6 Astra or OPS but at the 80% to 90% level.
In the brain tumor identification task, the model's judgments of the tumor type in all six brain scan images were wrong.
The total cost of all test tasks was only 6 USD (the UP主 topped up 50 USD and had 44 USD remaining).
All Facts (8)
The model has been open-sourced on Hugging Face and ModelScope, totaling about 510GB, and the community has already released quantized versions (e.g., Unsloth's 4bit GGUF at only 233GB).
50%UnverifiedOn benchmarks for agentic coding, general agents, knowledge, automation and others, DeepSeek C.E Flash 1017 improves by several percentage points over the previous generation, beats open-source models such as GLM5.3 and Kimi K3, and performs almost on par with GPT-5.6 Sol.
50%UnverifiedOn KernelBench (low-level GPU code writing), DeepSeek C.E Flash 1017 ranks third, behind only the latest Claude model, but beats Gemini 4 Argon, Claude Fable and GPT Leo Astra.
50%UnverifiedIn the music creation test, the model's first-round result was poor, and the second-round melody was similar to other models on the market; the UP主 considered it inferior to GPT-6 Astra or OPS but at the 80% to 90% level.
50%UnverifiedThe total cost of all test tasks was only 6 USD (the UP主 topped up 50 USD and had 44 USD remaining).
50%UnverifiedIn the brain tumor identification task, the model's judgments of the tumor type in all six brain scan images were wrong.
50%UnverifiedIn the UP主's hands-on test, DeepSeek C.E Flash 1017 ran for 36 minutes to complete the task of explaining the measurement of Earth's circumference, consuming about 22 million tokens with a 99% cache hit rate.
50%UnverifiedIn the Unreal Engine procedural 3D game task, the model ran 16 rounds, and the critic agent gave it a maximum score of 5.5 out of 10, falling short of the 8.5 AAA standard, consuming about 375 million tokens with a 99.5% cache hit rate.
50%