DeepSeek 4.1 Flash

DeepSeek open-sourced the inference framework DeepInfer on October 20, providing optimized operators, multi-card parallelism and asynchronous scheduling capabilities for V4.1 Flash, under the Apache 2.0 license.

DeepSeek has released an experimental multimodal model, DeepSeek V4 Flash Vision EXP, which performs excellently in both text and multimodal tasks.

DeepSeek 4.1 Flash and Gemini 4 Argon are two large models released in the same period.

DeepSeek 4.1 Flash's open-source weights can be downloaded on Hugging Face and ModelScope, the community has already released quantized versions, and the most compressed version is greatly reduced in size but still requires high-end GPUs.

DeepSeek 4.1 Flash consumed about 22 million tokens in the motion graphics explainer video task, with a 99% cache hit rate.

In the 3D game procedural generation task, DeepSeek 4.1 Flash ran 16 rounds of iteration and scored at most only 5.5 points (target 8.5 points), with overall intelligence at about 80% of frontier models.

In the Airbnb listing Blender reconstruction task, DeepSeek 4.1 Flash took about 1 hour 37 minutes, but its detail fidelity was insufficient (wrong bed shape, wrong lamp positions, etc.), and the evaluator believed it would be nearly impossible for other open-source models to complete.

In a six-question test on identifying tumors in brain scans, DeepSeek 4.1 Flash got all answers wrong; the evaluator noted that even the best models can only answer about one question correctly on this task.

DeepSeek 4.1 Flash is the first model to pass the 'find the frog' visual test, which closed-source models including the Opus and GPT series had previously failed to find.

On public benchmarks, DeepSeek 4.1 Flash exceeds its predecessor by several percentage points on most dimensions such as intelligent coding, common sense, and automation, leads open-source models such as GLM5.3 and Kimi K3, and is roughly on par with some top closed-source models; it ranks third on the GPU code benchmark.

The total cost of all test tasks was only 6 USD (the UP主 topped up 50 USD and had 44 USD remaining).

In the fourth task, complex scenes and NPC behavior planning, the DeepSeek 4.1 Flash version was rich in detail: cars waiting at red lights, sidewalks following traffic-light logic, NPCs that can be clicked to follow the view and each with different names, and relatively high modeling quality, with only the streetlights slightly tilted.

In the fifth task, a 15-second 3D text animation, the creator said he preferred the modeling texture of DeepSeek 4.1 Flash, while Gemini 4 Argon performed better on musical rhythm and animation pacing.

Of the six tasks, DeepSeek 4.1 Flash and Gemini 4 Argon each won three; DeepSeek was superior in Subway Surfers, 3D world generation and complex scene building, while Gemini performed better in the boss battle, text animation pacing and storyboard narrative animation.

In the third task, generating an explorable 3D world from a photo, DeepSeek 4.1 Flash had high fidelity, with excellent maple leaf paper-cutout modeling and falling-leaf texture, but building interiors were empty shells.

In the first task, a 3D Subway Surfers clone, the version generated by DeepSeek 4.1 Flash ran smoothly, with jumping, left-right movement, sliding and other operations free of bugs, and building modeling detail on par with the previous generation V4 Pro.

In the second task, a 3D boss battle game, the DeepSeek 4.1 Flash version had a minor bug where the character automatically rolled back after rolling, the boss modeling resembled the style of "Guobao Te Gong", and the second-phase skills triggered normally.

In the chess game comparison, the version generated by Opus 5.5 was clearly superior to Gemini 4 Argon's version in animation smoothness, scene lighting, and sound effect coordination, and could automatically generate cinematic replays of the match's highlight moments.

Among the 38 free models, only 4 are suitable for writing code: GLM5.3, DeepSeek 4.1 Flash, Nemotron 3 Ultra, and Poolside Laguna XS 2.1.

DeepSeek officially stated that 4.1 Flash adopts a new architectural design, making inference faster and more efficient.