DeepSeek V4 Flash
DeepSeek V4 Flash is a lightweight language model released by DeepSeek and a streamlined variant of the DeepSeek V4 series. While maintaining strong language understanding and generation capabilities, the model is optimized for inference efficiency and supports Multi-Token Prediction (MTP) speculative decoding technology to accelerate the inference process. It is suited for application scenarios with high demands on response speed and computational resources.
Core Facts
DeepSeek V4 Flash is an MoE multimodal model with a total parameter count of 284B.
DeepSeek V4 Flash is a cost-effective inference model released by DeepSeek, optimized for speed and cost.
DeepSeek V4 Pro采用了MoE（Mixture of Experts，混合专家）架构，在推理时只激活部分参数以降低计算成本
Timeline (last 90 days)
Based on public data for high power plus MLX or Ollama, the decoding speed of different models on M4 Max is roughly a ladder as follows: Nemotron 3 Nano about 160 tokens/s, Qwen 3.5 35B at 107 to 130 tokens/s, Gemma 4 26B about 75 tokens/s, DeepSeek V4 Flash about 27 tokens/s, and dense 27B about 22 tokens/s.
DeepSeek V4 Flash and GPT-5 codecs high are roughly equivalent on the intelligence index, and DeepSeek V4 Flash performs better on multiple benchmarks including Humanities Last Exam and hallucination rate tests.
DeepSeek V4 Flash's input tokens are 21 times cheaper and output tokens 55 times cheaper; even assuming 3 times lower efficiency, the net calculation still yields approximately a 20-fold improvement.
The team overrode the Haiku slot with DeepSeek V4 Flash through local configuration, having it handle templated scripting tasks such as Remotion video scripts, subtitle exports, and screenshot packaging.
The vast majority of coding for the HoopMind (NBA Chat) project was completed by Claude Code as the primary agent, supplemented by DeepSeek V4 Flash as a sub-agent.
The author recommends using DeepSeek V4 Flash for vulnerability discovery and using stronger models such as DeepSeek V4 Pro or Opus for validation.
Colibri provides unified access via coli chat, coli serve, and coli web, supporting models including GLM 5.2/5.3, Kimi K3, DeepSeek V4 Flash, and Qwen3.
Real-world testing with the 284-billion-parameter DeepSeek V4 Flash shows token generation rising from 37 to 53 tokens/s (approximately 1.5x), and Prompt Processing jumping from 483 to 1,485 tokens/s (approximately three times).
DeepSeek V4 Flash failed 13 out of 22 multi-turn tests (fail_fast 45%), and Gemini also failed 13.
Community benchmarks show that the Pro plan can process approximately 10 billion tokens per month running DeepSeek V4 Flash, but only about 1.6 billion tokens running GLM 5.1, a difference of 6x.
40 more timeline events
All Facts (20)
DeepSeek V4 Flash is an MoE multimodal model with a total parameter count of 284B.
80%VerifiedDeepSeek V4 Flash is a cost-effective inference model released by DeepSeek, optimized for speed and cost.
80%VerifiedDeepSeek V4 Pro采用了MoE（Mixture of Experts，混合专家）架构，在推理时只激活部分参数以降低计算成本
80%VerifiedDeepSeek is deploying native multimodal capabilities at the Flash level rather than the flagship version first, with the intent to prioritize capturing high-frequency, large-scale application scenarios.
75%VerifiedDeepSeek V4.1 Flash scores lower than the previous version on GPQA, HLE, and other benchmarks.
75%VerifiedCompared to the previous-generation V4 Flash, V4.1 Flash reduces the demand for high-bandwidth memory (HBM) to one-quarter and the demand for SSD persistent storage to one-eighth.
75%VerifiedV4 Flash's cache size is only 7% of that of V3.2.
70%VerifiedV4.1 Flash surpassed Opus 5 on the DeepSWE benchmark by a narrow margin of 0.2 points.
70%VerifiedThe security advantage of local deployment lies in keeping data on the device, eliminating the risk of man-in-the-middle attacks during network transmission, and reducing reliance on trust in cloud service providers.
65%VerifiedDeepSeek has launched the experimental multimodal model V4-Flash-Vision-EXP on its API platform, with text capabilities on par with V4 Flash.
90%VerifiedDeepSeek has released an experimental multimodal model, DeepSeek V4 Flash Vision EXP, which performs excellently in both text and multimodal tasks.
90%VerifiedDeepSeek has launched the experimental multimodal model V4-Flash-Vision-EXP on its API platform, with text capabilities on par with V4 Flash and multimodal agent benchmark performance close to Opus 4.8 levels.
85%VerifiedDeepSeek officially launched DeepSeek Harness (abbreviated as DSH) following the release of the DeepSeek V4 Pro model.
70%UnverifiedV4 Pro的代码生成速度明显慢于V4 Flash
90%Partially VerifiedOn a dual-Spark system, DeepSeek V4 Flash (FP8 quantization) peaks at approximately 53 tokens per second.
65%UnverifiedEwan uses the DeepSeek V4 Flash model for most of his work, describing it as very fast, relatively inexpensive, and highly capable.
60%UnverifiedIn this test, the model's output speed in short-context scenarios is approximately 20 tokens/second, dropping to approximately 13 tokens/second for contexts of 200K+.
60%UnverifiedDeepSeek V4 Flash is directly compatible with Codex and can be connected without CC Switch.
60%UnverifiedThe smartest models tend to be the most token-efficient, because they can more precisely understand intent, organize output more efficiently, and avoid redundant trial and error.
60%UnverifiedThe UD-Q8_K_XL quantized version of DeepSeek V4 Flash has a file size of 162 GB, uses MXFP4+BF16 mixed precision, and achieves completely lossless quality.
60%