DeepSeek V4.1 Flash shipped on September 10, 2026, and the spec sheet reads like a typo: a 552-billion-parameter MoE backbone with a 196-billion-parameter Engram conditional-memory module attached. Despite the "Flash" name, DeepSeek's own evaluations show it beating the much larger V4 Pro on performance, cost, speed, and total task time, and third-party indexes (Artificial Analysis) score it above V4-Pro as well.
Two transition details matter for anyone building on the DeepSeek API: from September 14, requests to deepseek-v4-pro are served by V4.1 Flash while V4.1 Pro is in development, and the legacy deepseek-v4-flash model name has been retired - those requests are now served by V4.1 Flash at Flash pricing.
The "Flash" label undersells it. This is not a small model that got cheap; it is a full-size model that learned to spend compute only where it pays.
CED: an asymmetric encoder-decoder that halves prefill
V4.1 Flash replaces the pure decoder-only stack with a Causal Encoder-Decoder (CED): 40 transformer layers split into a 20-layer causal encoder and a 20-layer decoder. Prompt tokens only traverse the encoder half. The decoder does not recompute its own global KV cache layer by layer - it projects it directly from the encoder's final hidden states.
The activation asymmetry is the point: 8B parameters per token during prefill, 16B during decode. Agent workloads are prefill-heavy - every tool call replays the whole conversation context - so cheaper prefill is exactly where agent economics improve.
CSA2 and the 890-byte KV cache
The attention stack is CSA2 (Compressed Sparse Attention v2). Each layer runs in one of three static modes - Full, Reindex, or Reuse - sharing the main KV and indexer keys across layers instead of every layer recomputing everything. The main KV cache is stored in FP4 (E2M1), bringing the global cache down to roughly 890 bytes per token: about a quarter of V4-Flash, and roughly 437x smaller than DeepSeek-V1. Persistent SSD storage drops to about an eighth.
The practical consequence: decode compute grows only ~25% when context scales from 4K to 1M tokens. Long context stops being a luxury and becomes a default operating mode.
Engram: a 196B lookup dictionary for facts
The least-discussed piece may matter the most. Engram is a conditional-memory module accessed by token lookup rather than by keeping parameters hot - a third memory mechanism alongside attention and pretrained weights. It comes out of the January 2026 DeepSeek and Peking University paper Conditional Memory via Scalable Lookup.
The economic implication is the real story: historically, adding knowledge capacity meant adding active compute and cost with it. Engram decouples the two - a 196B-parameter table that costs almost nothing per token because it is only consulted sparsely. V4.1 Flash is the first production proof that bigger does not have to mean more expensive.
Reasoning effort: a dial you should actually use
The API exposes reasoning effort in three tiers (low / high / max). Official benchmarks climb with effort, but so does token burn: the top tier costs several times the output tokens of the middle tiers for a marginal accuracy gain.
Community testing also points to a change in thinking style: where R1-style reasoning explored (backtrack, reconsider, start over), V4.1 tends to verify - it checks rather than guesses. Verification is thorough, which means the model can over-think easy tasks. Practical rule: max for hard agentic and debugging work, low for routine generation and extraction.
Pricing: off-peak vs peak, and the 1/50 cache hit
| Token type | Off-peak (¥/M tokens) | Peak (¥/M tokens) |
|---|---|---|
| Input - cache hit | ¥0.02 | ¥0.04 |
| Input - cache miss | ¥1.00 | ¥2.00 |
| Output | ¥4.00 | ¥8.00 |
Peak hours are weekdays 9:00-12:00 and 14:00-18:00 Beijing time; everything else - including all weekend - runs at off-peak rates.
The cache-hit line is the sleeper hit: a repeated token costs 1/50 of a fresh one. Keep system prompts, tool schemas, and long conversation histories byte-stable at the start of your prompts, and an agent that replays the same context hundreds of times pays almost nothing for it. If you inject a timestamp at the front or reorder sections between calls, you nuke your own cache - structure prompts so the stable parts come first.
What this unlocks for agent builders
- Full-repository context: a 1M-token window backed by an 890-byte/token cache makes "throw the whole codebase in" viable instead of heroic.
- Log and document analysis: entire runs, manuals, or knowledge bases can stay resident without the bill exploding.
- Long agent loops: cheap prefill plus cheap cache hits remove the two costs that used to punish multi-hour tool-calling chains.
- A note on benchmarks: numbers like DeepSWE 74.2 and Terminal-Bench 2.1 90.6 are publisher (DeepSeek) scores at max effort - treat them as a strong signal, and wait for independent replication before rewriting your eval suite.
Trying V4.1 Flash from outside China
The friction for many international developers is not the model - it is access: sign-up and payment on the official platform typically require Chinese payment methods that most foreign cards cannot complete.
AiCredits sells DeepSeek API credits payable with an international credit card. You pick a plan, and a working key for the OpenAI-compatible endpoint is shown on the success page and emailed within seconds - deepseek-flash (V4.1 Flash) is the current flagship model.
curl https://aicreditsapi.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer sk-your-aicredits-key" \
-d '{
"model": "deepseek-flash",
"messages": [{"role": "user", "content": "Summarize this repository README"}]
}'
FAQ
Is DeepSeek V4 Pro discontinued? No. DeepSeek announced V4 Pro remains available after September 14 with billing unchanged; during the transition to V4.1 Pro, requests to it may be served by V4.1 Flash infrastructure.
What happened to the deepseek-v4-flash model name? It is retired. Requests to the old name are served by V4.1 Flash and billed at Flash prices - no code changes required on your side.
How is the cache-hit price applied? When the prefix of your prompt matches a recent request, those input tokens bill at ¥0.02/M (off-peak) instead of ¥1/M. Prefix stability is what triggers it.
Can I use DeepSeek V4.1 Flash without a Chinese phone number or payment method? Yes - AiCredits offers credits payable by international credit card, with instant key delivery.
Top comments (0)