DeepSeek V4.1 Flash Released: What Changed, What It Costs

On September 10, 2026, DeepSeek released DeepSeek-V4.1-Flash. This isn't a minor bump to the V4-Flash line — it's the first model of a new architecture family, with an asymmetric design that uses different numbers of active parameters for input and output, native visual understanding, and a KV cache one quarter the size of the previous generation. The API model name is deepseek-flash, live on the DeepSeek API as of launch day.

You can run DeepSeek V4.1 Flash in meshcode right now — the catalog was updated to the new name and rates within days of release. How to find it in the app is covered below.

The new architecture: 552B MoE, 8B/16B active

Per the official announcement, V4.1 Flash's structure looks like this:

• 552B-parameter MoE — the total is large, but only a tiny fraction of parameters activates per token.
• Causal Encoder–Decoder architecture — 8B active parameters for input processing, 16B for output generation. This asymmetric design budgets reading and writing separately, optimizing the cost of consuming long contexts independently from the cost of generating tokens.
• Native multimodality — one model understands images, no separate vision variant. What was split across models in the V4 generation (a separate Vision-Exp release) is now unified.
• 1M context, up to 384K output — both thinking and non-thinking modes are supported, with thinking as the default.

DeepSeek states that new pre-training methods plus larger-scale RL post-training put benchmark results ahead of its flagship models, including V4-Pro. These are the manufacturer's own launch claims, and meshcode deliberately assigns no capability rankings to models — the orchestrator is designed to judge fitness from actual delegation history and your own ordering, not from a leaderboard we invent.

KV cache at one quarter: this hits agent costs directly

Agent coding workloads re-read the same context over and over: a long system prompt, a codebase snapshot, the results of previous tool calls — all resent on every request. Whether those hit the cache is often the dominant cost factor, and DeepSeek's cache-hit rate runs about 1/50th of the cache-miss rate, so cache charges can make up the bulk of an agent's bill.

DeepSeek reports that V4.1 Flash needs just one quarter of the HBM and one eighth of the SSD storage for its KV cache versus the previous generation. That's more than a server-cost saving — it structurally lowers the total cost of cache-heavy agent workloads, which is exactly how DeepSeek justifies the lower prices that come with it.

Pricing: peak vs. off-peak is a 2x spread

New rates took effect on September 10 alongside the model. Per 1M tokens (official pricing page, retrieved 2026-09-14):

Off-peak is exactly 50% of peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday — everything else, including all weekend hours, is off-peak. Moving flexible work — batch refactors, test generation, documentation sweeps — into off-peak windows halves the token bill on identical output.

Note the $0.006 cache-hit input rate at peak, too. In long agent sessions where most input hits the cache, the effective input cost lands far below the headline rate.

What happens to the old model names

Worth knowing if you have DeepSeek model IDs in your configs (official docs, retrieved 2026-09-14):

• deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired. The names still work for compatibility, but requests are served by V4.1 Flash and billed at V4.1 Flash rates.
• deepseek-v4-pro keeps running. It was originally slated to route to V4.1 Flash starting September 14, but DeepSeek has since announced — in response to user demand — that V4-Pro API service continues with billing unchanged (further notice if that changes).

New projects should just use deepseek-flash. Existing code calling deepseek-v4-flash needs no changes and is already getting V4.1 Flash. V4.1-Pro is announced as coming later.

Open weights and self-hosting

V4.1 Flash ships with open weights on Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash), along with a technical report. DeepSeek says it will work with the open-source community on inference support and explore more deployment options. For self-hosters, the KV cache reduction is a memory story: the same context length should fit on meaningfully less hardware than the previous generation.

Running V4.1 Flash in meshcode

DeepSeek V4.1 Flash is supported in meshcode today. Account → Models in the app shows the live catalog and current rates — since both model lists and prices keep moving, in-app is always the most accurate view.

Two ways to use it:

• In meshcode's own model stack — pay-as-you-go credits (top up from $1), billed per token. The natural pattern: route bulk work to the low-cost model and reserve expensive models for judgment calls.
• In parallel panes — run DeepSeek on a refactor in one pane while a different model reviews in another. You're not choosing one model; you're splitting work across them.

For how V4-generation economics compare against Claude Code's flat rate, see DeepSeek V4 vs Claude Code: Which Is Cheaper?; for the pay-as-you-go vs. subscription math itself, AI Coding Agent Token Costs in 2026.

The short version

Three things matter for coding workloads. First, the asymmetric active-parameter split (8B input / 16B output) marks a new architecture family that budgets reading and writing costs separately. Second, the 4x KV cache reduction changes the cost structure of exactly the workloads agents run — long, cache-heavy sessions. Third, the 50% off-peak discount means scheduling alone can halve what you pay.

Rates and model names keep changing — before any big decision, trust the official pricing page and meshcode's in-app Models list over the table above.

👉 Download meshcode — Mac · Windows