DeepSeek-V4.1-Flash
A large mixture-of-experts language model released by DeepSeek, employing a 552B MoE architecture. It supports visual input and a one-million-token context window, and introduces the Engram memory architecture along with a compressed KV sharing mechanism, optimized for agentic use cases.
Core Facts
Pangu 2.0 Flash adopts a MoE (Mixture of Experts) architecture with a parameter scale of 92B A6B (92B total parameters, 6B activated parameters).
DeepSeek-V4.1-Flash's KV cache is compressed to 890 bytes per token, only one-quarter that of V4 Flash.
Timeline (last 90 days)
Earlier tests by Artificial Analysis and Vals.ai showed that V4.1 Flash ranks among the top open-weight models but lags on SRE-type tasks and legal-agent benchmarks.
The top of terminal-bench-2.0 is mindlab-research/Macaron-V1-Venti (87.6), terminal-bench-2.1 belongs to DeepSeek-V4.1-Flash (90.6), and the base version terminal-bench is led by Ling-3.1-flash with 46.03.
DeepSeek-V4.1-Flash-1017 is a post-training checkpoint, not a new architecture, and is still based on V4.1-Flash released in September.
Among six zeroth-order cases on DeepSeek-V4.1-Flash, only one new target token promotion occurred, along with one regression.
The MoE structure of DeepSeek-V4.1-Flash has 40 layers, each layer has 384 routed experts, and each token activates only 6 experts per layer.
DeepSeek-V4.1-Flash has 763B parameters, BF16 weights are about 1.5 TB, the official FP4/FP8 checkpoint is 475.2 GiB, and coolbho3k's 3-bit EXL3 package compresses it to 309.3 GiB.
One change in v0.32.0rc2 is to pin the evaluation of DeepSeek-V4.1-Flash-NVFP4 to its FP8-Engram version.
When deploying DeepSeek-V4.1-Flash or Moonshot Kimi K2.5 on clusters with insufficient specifications, vLLM either OOMs during the weight initialization stage or evicts KV cache blocks on every concurrent multi-turn Agent request.
Models using MLA such as DeepSeek-V4.1-Flash compress the KV state to about 2.4 GB per 32k stream (FP8 KV), while wider architectures consume 3.0 GB–3.8 GB per 32k stream.
vLLM's three optimizations for DeepSeek-V4.1-Flash (SWA bounded replay, MegaAttention and Mega-Gate kernels, operator fusion and CUDA graphs) achieve 5.3x throughput under a 150 TPS constraint, a 1.9x speedup at low concurrency, and about a 70% reduction in TTFT at 100K throughput.
37 more timeline events
All Facts (20)
Pangu 2.0 Flash adopts a MoE (Mixture of Experts) architecture with a parameter scale of 92B A6B (92B total parameters, 6B activated parameters).
80%VerifiedDeepSeek-V4.1-Flash's KV cache is compressed to 890 bytes per token, only one-quarter that of V4 Flash.
65%UnverifiedThe release party claimed that the cost of DeepSeek-V4.1-Flash is only 1/40 that of Opus 5 and GPT-5.6 Sol.
70%UnverifiedDeepSeek-V4-Flash-Vision-Exp adopts a MoE architecture with approximately 305B total parameters, activating only approximately 13B per token, for an activation rate of approximately 4.6%.
60%UnverifiedEarlier tests by Artificial Analysis and Vals.ai showed that V4.1 Flash ranks among the top open-weight models but lags on SRE-type tasks and legal-agent benchmarks.
50%UnverifiedThe top of terminal-bench-2.0 is mindlab-research/Macaron-V1-Venti (87.6), terminal-bench-2.1 belongs to DeepSeek-V4.1-Flash (90.6), and the base version terminal-bench is led by Ling-3.1-flash with 46.03.
50%UnverifiedDeepSeek-V4.1-Flash-1017 is a post-training checkpoint, not a new architecture, and is still based on V4.1-Flash released in September.
50%UnverifiedAmong six zeroth-order cases on DeepSeek-V4.1-Flash, only one new target token promotion occurred, along with one regression.
50%UnverifiedThe MoE structure of DeepSeek-V4.1-Flash has 40 layers, each layer has 384 routed experts, and each token activates only 6 experts per layer.
50%UnverifiedDeepSeek-V4.1-Flash has 763B parameters, BF16 weights are about 1.5 TB, the official FP4/FP8 checkpoint is 475.2 GiB, and coolbho3k's 3-bit EXL3 package compresses it to 309.3 GiB.
50%UnverifiedOne change in v0.32.0rc2 is to pin the evaluation of DeepSeek-V4.1-Flash-NVFP4 to its FP8-Engram version.
50%UnverifiedWhen deploying DeepSeek-V4.1-Flash or Moonshot Kimi K2.5 on clusters with insufficient specifications, vLLM either OOMs during the weight initialization stage or evicts KV cache blocks on every concurrent multi-turn Agent request.
50%UnverifiedModels using MLA such as DeepSeek-V4.1-Flash compress the KV state to about 2.4 GB per 32k stream (FP8 KV), while wider architectures consume 3.0 GB–3.8 GB per 32k stream.
50%UnverifiedvLLM's three optimizations for DeepSeek-V4.1-Flash (SWA bounded replay, MegaAttention and Mega-Gate kernels, operator fusion and CUDA graphs) achieve 5.3x throughput under a 150 TPS constraint, a 1.9x speedup at low concurrency, and about a 70% reduction in TTFT at 100K throughput.
50%UnverifiedDeepSeek-V4.1-Flash is based on multimodal pre-training with 45T tokens, supports context reasoning intensity continuously adjustable from 1 to 100, and uses the MIT license.
50%Unverifieddeepseek-v4.1-flash belongs to the lightweight, high-throughput flash-class models, which are typically lower in price and called frequently.
50%UnverifiedThe BF16 weight files for Naive-N0.5-Flash are approximately 618GB, and the FP8 format version weights are approximately 315GB
50%UnverifiedDeepSeek-V4.1-Flash is open-sourced under the MIT license, natively supports multimodal vision, and has a context length of 1 million tokens.
50%UnverifiedThe read-light, write-heavy dynamic activation pattern of DeepSeek-V4.1-Flash is suited for agentic workflows that require frequent reading of large amounts of context with intermittent output.
50%UnverifiedvLLM provided native support for DeepSeek-V4.1-Flash from the day of its release and completed validation on both NVIDIA and AMD GPU platforms.
50%