DeepSeek Releases V4.1 Flash Slashing Cached Token Costs by 60 Percent
DeepSeek launched V4.1 Flash, a 552-billion parameter model that activates only 8 billion parameters on input and 16 billion on output. The architecture shrinks Key-Value cache sizes to 890 bytes per token, lowering cached input rates to $0.003 per million tokens while scoring 74.2 on DeepSWE v1.1.
Impact: High
Why it matters
You can cut repetitive context costs by 60% in multi-turn coding loops without modifying your API client, as legacy flash model endpoints automatically route to V4.1 Flash.
TL;DR
- 01Cached input rates drop by 60 percent down to $0.003 per million tokens off-peak.
- 02KV cache footprint drops to 890 bytes per token via FP4 quantization and Compressed Sparse Attention 2.
- 03Context window supports 1M input tokens and up to 384k output tokens with native vision by default.
- 04Matches Claude Opus 5 on DeepSWE v1.1 (74.2 vs 74.0), but trails significantly on Terminal-Bench 3.0 (30.0 vs 43.3).
Key facts
- Total / Active Parameters
- 552B total, 8B active input, 16B active output
- Cached Input Cost (Off-peak)
- $0.003 per 1M tokens (-60%)
- KV Cache Footprint
- 890 bytes per token
- Context Window Limits
- 1M input, 384k output tokens
- DeepSWE v1.1 Score
- 74.2 (self-reported)
- Terminal-Bench 3.0 Score
- 30.0 (self-reported)
Causal Encoder-Decoder Architecture
DeepSeek V4.1 Flash transitions away from standard monolithic dense execution by using a Causal Encoder-Decoder structure with 552 billion total parameters distributed across 20 encoder and 20 decoder layers. Input processing activates only 8 billion parameters, while output token generation activates 16 billion parameters. This structural sparsity dramatically reduces compute load during prompt ingestion.
KV Cache Compression and Pricing
Persistent memory footprint is the primary bottleneck for massive context windows. V4.1 Flash implements Compressed Sparse Attention 2 and FP4 cache quantization, cutting Key-Value (KV) cache consumption to 890 bytes per token (one-fourth of V4-Flash and one-eighth of prior baselines).
- Cached Input: $0.003 per million tokens off-peak (60% discount).
- Uncached Input: 33% price reduction.
- Output Tokens: 11% price reduction.
- Context Limits: 1,000,000 total tokens with up to 384,000 output tokens.
Migration and Vision Integration
Multi-modal vision processing is now baked directly into the core pre-training via DeepSeek-ViT, eliminating the previous deepseek-v4-flash-vision-exp split. Existing callers targeting deepseek-v4-flash are already routed to V4.1 Flash, and deepseek-v4-pro endpoints transition on September 14 at Flash pricing tiers.
Try it in 2 minutes
curl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $DEEPSEEK_API_KEY" \
-d '{
"model": "deepseek-v4-flash",
"messages": [{"role": "user", "content": "Analyze repo AST"}],
"max_tokens": 4096
}'bash
✓ When to use
- Large-context codebase ingestion and documentation indexing where prompt caching reduces operating costs.
- Autonomous code generation and refactoring pipelines evaluated on SWE-bench and DeepSWE benchmarks.
- Unified multimodal image and code parsing workflows requiring high token throughput at low cost.
✕ When NOT to use
- Interactive agent tasks that require extended shell navigation and live terminal debugging (Terminal-Bench score 30.0 vs Opus 5 43.3).
- Strict low-latency edge deployments requiring full offline execution on modest local GPU memory.
What to do today
- Update API routing configs to leverage cached inputs during repetitive code context injection.
- Retire dedicated vision-exp endpoints since multimodal processing is now built into deepseek-v4-flash.
- Evaluate agent tool loops using Terminal-Bench to check whether the 30.0 shell score impacts your workflow.
Sources