552B MoE
8B active parameters per input token and 16B during output generation
Released September 10, 2026
DeepSeek V4.1 Flash: API, Architecture, Benchmarks & Harness Guide
DeepSeek V4.1 Flash is a 552B-parameter multimodal MoE model built for long, tool-heavy Agent workloads. The official API uses the model ID deepseek-flash, supports text and images, and exposes a context window of up to one million tokens.
- Author
- DeepSeekAgent.io Editorial Team
- Published
- Updated
What changed in V4.1 Flash
1M context
Long sessions, large repositories, images, and tool histories can share one context
890 bytes/token KV
DeepSeek reports roughly one quarter of V4 Flash's global KV-cache footprint
Native multimodal
Text and image inputs are handled by the same model and prompt format
API migration: use deepseek-flash
The stable API-facing name is deepseek-flash. DeepSeek says the retired deepseek-v4-flash and deepseek-v4-flash-vision-exp IDs temporarily route to V4.1 Flash for compatibility. New configurations should use the current ID so the intended target remains explicit.
| Status | Model ID | Meaning |
|---|---|---|
| Recommended | deepseek-flash | Current V4.1 Flash route |
| Legacy alias | deepseek-v4-flash | Temporarily routes to V4.1 Flash |
| Legacy vision alias | deepseek-v4-flash-vision-exp | Temporarily routes to V4.1 Flash |
Why the architecture matters to an Agent Harness
The new Causal Encoder–Decoder design separates a 20-layer causal encoder from a 20-layer decoder. CSA2 shares and reuses sparse-attention state across layers, while FP4 caching reduces the global KV footprint. These choices target the expensive shape of Agent work: large prompts and histories, repeated cache reads, tool calls, and comparatively smaller generated actions.
Official Agent benchmark snapshot
The numbers below are reported by DeepSeek at maximum reasoning effort with temperature 1.0 and top_p 0.95. Code-agent evaluations use a one-million-token context. They are useful release evidence, not a guarantee for every repository or production workflow.
| Benchmark | V4 Flash | V4.1 Flash | Change |
|---|---|---|---|
| Terminal-Bench 2.1 | 82.7 | 90.6 | +7.9 |
| Terminal-Bench 3.0 | 7.6 | 30.0 | +22.4 |
| Terminal-Bench 4.0 | 7.0 | 31.2 | +24.2 |
| DeepSWE v1.1 | 54.4 | 74.2 | +19.8 |
| NL2Repo-Bench | 54.2 | 64.0 | +9.8 |
The Harness angle
DeepSeek evaluated V4.1 Flash across Claude Code, Codex, OpenCode, Pi, mini-SWE, DSH Minimal, DSH Standard, and DSH PTC. That table is especially useful because it holds the model constant while changing the scaffold. On DeepSWE v1.1, mini-SWE reports 74.2 and DSH Minimal 72.6; on Terminal-Bench 2.1, DSH Minimal reports 90.6. The result supports testing profiles independently instead of treating all DSH modes as one system.
Continue with the V4.1 Harness series
Harness setup and model ID
Configure deepseek-flash, images, reasoning effort, and migration checks.
Read guide →DeepSWE reproduction
Follow the official Pier and DSH Minimal methodology and audit the outputs.
Read guide →V4.1 Flash vs V4 Flash
Separate official evidence from an early community quick test.
Read guide →Frequently asked questions
What model ID should I use for DeepSeek V4.1 Flash?
Use deepseek-flash for new DeepSeek API and DeepSeek Harness configurations.
Does V4.1 Flash accept images?
Yes. DeepSeek describes it as natively multimodal, and the current DSH catalog declares text and image input modalities for deepseek-flash.
Does the 1M context mean every long Agent run is reliable?
No. It is a capacity limit. Long-run quality still depends on prompt structure, tool feedback, compaction, reasoning effort, and workload-specific testing.