DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash tops Terminal-Bench 2.1 and DeepSWE against Opus 5.0 and GPT-5.6, activates 8 billion parameters to do it, and ships under MIT. It also scores 30.0 where Opus scores 43.3 on the next version of that same terminal benchmark. Both of those facts are on DeepSeek's own model card, and reading them together is more useful than reading either alone.

552B on the disk, 8B in the math

The architecture is the headline. DeepSeek calls it a Causal Encoder-Decoder: 40 layers split into a 20-layer encoder and a 20-layer decoder, with Mixture-of-Experts layers carrying one shared expert and 384 routed experts, six of which fire per token. The backbone is 552B parameters. Activated parameters are 8B during prefill and 16B during decode.

Alongside that sits a 196B block DeepSeek calls Engram conditional memory, accessed sparsely. Total download is about 763B parameters' worth of weights in safetensors, with BF16, F32, FP8 E4M3 and INT8 tensor types present.

The number that will matter most in production is the KV cache: 890 bytes per token, roughly a quarter of DeepSeek-V4-Flash and, by DeepSeek's count, about 437 times smaller than DeepSeek-V1. Context runs to 1M tokens, with sparse attention trained at 64K and extended to 1M over 34T tokens. If you have ever abandoned a long-context deployment because the cache would not fit, that line is the release.

Vision is native this time. A DeepSeek-ViT encoder trained from scratch with 2D-RoPE and 3x3 pixel-unshuffle downsampling feeds a two-layer MLP projector, so images become visual embeddings the model consumes directly. V4-Flash had no multimodal row at all.

Where it wins, and where it does not

These are DeepSeek-reported, run with the Minimal mode of their own DeepSeek Harness at a 1M-token context, temperature 1.0 and top_p 0.95. A lab evaluating its model inside its own harness is a caveat worth carrying through the whole table.

The top four rows are a genuine result. An open MIT model with 8B active parameters edging Opus 5.0 on DeepSWE and taking Terminal-Bench 2.1 outright is not a rounding error, and the Codeforces rating of 3471 is the highest in DeepSeek's comparison set.

The bottom four rows are the correction. On Terminal-Bench 3.0 and 4.0, Opus is thirteen and twenty points ahead. On Humanity's Last Exam the gap is nearly twenty points. The pattern is consistent: V4.1-Flash is at or above frontier on the benchmarks that have been public long enough to optimise against, and clearly behind on the newest and hardest ones. Note that it still beats K3 handily on 3.0 and 4.0, so this is not a weak model. It is a model whose ranking depends a lot on which benchmark generation you pick.

One more caution on reading the card: the widely quoted HumanEval 79.4 and GSM8K 93.0 come from the base-model pretraining table, not the instruct evaluations. They are few-shot completions on the base checkpoint and say nothing about the agentic model people will actually deploy.

MIT, and that still matters

DeepSeek ships this under MIT, as it did V4-Flash. No attribution string, no revenue ceiling, no research-only clause, no separate commercial license to negotiate. Set that against the rest of the month: YuE2 went non-commercial, MiniMax Music 3 permits commercial use with a UI attribution and a $20M ceiling, and Qwen-Image-2.1 is research-only. DeepSeek remains the lab that just gives you the weights.

That is the real reason to care about this release even if the benchmark story is mixed. A 1M-context multimodal agentic model with 890-byte-per-token KV cache, under a license with no conditions, is a foundation you can build a product on without a lawyer.

Running it

This is not a laptop model. 763B parameters of weights means serious hardware even with 8B active, and the supported paths are server runtimes: vLLM, SGLang, Docker Model Runner and Transformers, with Colab and Kaggle notebooks available for poking at it.

The feature worth trying first is continuously controllable reasoning effort, exposed on a 1 to 100 scale. Most models give you a binary thinking toggle; a dial lets you tune spend against latency per request rather than per deployment.

If you want the practical comparison of serving stacks before you commit, our Ollama vs llama.cpp vs vLLM guide covers the tradeoffs, and for this model vLLM or SGLang are the realistic options.

Who should care

If you run coding agents and you have the hardware, this is the strongest MIT-licensed option available and the KV cache reduction changes what a long-context agent costs to keep alive. If you are picking a model for hard novel tasks rather than well-trodden benchmarks, the Terminal-Bench 3.0 and 4.0 rows say to keep testing Opus alongside it.

And if you do not have a multi-GPU server, this release is not for you today. It is worth watching anyway, because DeepSeek's Flash line is where the architecture experiments land first and the smaller distillations tend to follow.

Sources and further reading

• DeepSeek-V4.1-Flash model card and full benchmark tables
• DeepSeek on Hugging Face
• vLLM, one of the supported serving runtimes

Ten minutes: open the model card's agentic table and compare the Terminal-Bench 2.1 row against the 3.0 and 4.0 rows yourself. That single comparison is the most useful thing in this release, and it is a habit worth applying to every model card you read after it.

Tested on: not independently tested. Architecture, parameter counts, KV cache figures and every benchmark number are DeepSeek-reported, and the agentic evaluations were run inside DeepSeek's own harness. No independent replication exists at the time of writing. License and file formats were read from the Hugging Face model card on the date below.
Date checked: 2026-09-27