DeepSeek v4.1 Flash

TL;DR Highlight

DeepSeek V4.1 Flash shrinks KV cache to cut agent costs and adds native vision.

Who Should Read

An LLM API developer whose long-context coding agents are burdened by cache-hit costs.

Core Mechanics

Smallest model in the new architecture family
- It is a 552B-parameter MoE (mixture of experts) model.
- It has native visual understanding.
Only 8B active for input, 16B for output
- The new Causal Encoder–Decoder architecture activates 8B parameters for input and 16B for output.
- New pre-training methods and larger-scale RL post-training reportedly drive the benchmark gains.
KV cache is a fraction of the previous generation
- Compared with the previous generation, it needs 1/4 the HBM and 1/8 the SSD storage.
- DeepSeek says cache-hit charges are often a large share of agent costs, so compressing the cache cuts them.
Use it via deepseek-flash
- Set your model to deepseek-flash.
- V4-Flash and V4-Flash-Vision-Exp are retired, and the old names deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1-Flash.
Off-peak is 50% of peak

Community Reactions

Cache-hit price stood out
- One commenter noted cache hits cost $0.003 per million tokens.
- They shared a 447-turn coding task costing 55.057 on GPT-6-astra versus 0.615 on DeepSeek.
Questions about the price cut
- A commenter noted V4 flash was $0.10/$0.25 while V4.1 is $0.30/$1.20, asking how that is a cut.
- Another said it is about 3x pricier than V4 but faster and slightly more token efficient.
Only three reasoning levels in practice
- A commenter ran seven OpenRouter reasoning levels and found token counts did not line up with them.
- Official docs reportedly support low, high and max, with the others mapped onto those.
Benchmarks are not a blanket win
- One commenter noted that in DeepSeek's table it beats GPT-5.6 Sol on Terminal-Bench 2.1, DeepSWE and others but trails on GPQA Diamond and others.
- On a puzzle-game eval, a commenter found it burns more reasoning tokens than other models and Gemini 3.8 Flash did better.

How to Apply

Swap the model to deepseek-flash and compare
- Change your agent's model name to deepseek-flash and run a few identical tasks.
- Record quality and token usage side by side with your current model.
Find code using old model names
- Search your configs for deepseek-v4-flash and deepseek-v4-flash-vision-exp.
- Switch them to deepseek-flash before the temporary routing ends.
Move deferrable batch jobs off-peak
- Pick non-urgent jobs like nightly summaries or bulk classification.
- Schedule them in off-peak hours to pay 50% of peak rates.
Compute your cost split by cache share
- Pull input, output and cache-read tokens from a past agent run.
- Plug in different cache prices and see what share of total cost the cache takes.

Try It

copy & run

# Paste into Claude or ChatGPT (cost-structure calculator)

Here is the token usage of my agent task.
- Input tokens: {{input tokens}}
- Output tokens: {{output tokens}}
- Cache read tokens: {{cache read tokens}}

Using these prices (USD per million tokens), build a table of each cost and its share of the total.
- Model A: input {{A input price}}, output {{A output price}}, cache {{A cache price}}
- Model B: input 0.30, output 1.20, cache 0.003

Then tell me which model has the larger cache share and how many times the total cost changes if I switch.

Terminology

MoE (Mixture of Experts)An architecture that uses only some of many expert sub-networks per input. Like a big hospital where each patient sees only the relevant specialist.

KV cacheMemory that stores earlier tokens' computed results. It grows as the conversation gets longer.

HBMVery fast, expensive memory attached to the GPU. Large KV caches here drive up cost.

Active parametersThe parameters actually used to process one token. The smaller the number, the lighter the inference.

Cache hitReusing stored results for an already-processed prompt prefix instead of recomputing it.