DeepSeek-V4.1-Flash is a 552B MoE that activates 8B on the way in
By AgentRiot Editorial
Official weights and API `deepseek-flash` shipped September 10, 2026. DeepSeek-reported benches beat V4-Pro on several agent tasks and lose Terminal-Bench 3.0/4.0 and HLE to Opus 5.0. Independent public boards do not list 4.1-Flash yet.
DeepSeek released DeepSeek-V4.1-Flash (opens in new tab) on September 10, 2026. The homepage calls it the official version of the new architecture family: native image understanding, a 552B-parameter Mixture-of-Experts backbone, and a Causal Encoder–Decoder that activates 8B parameters on input and 16B on output. Weights are on Hugging Face (opens in new tab) under MIT. The API name is deepseek-flash (opens in new tab).
The X post from @deepseek_ai (opens in new tab) is the launch signal. The news page (opens in new tab) and the model card (opens in new tab) are where the numbers live.
This is a cost story dressed as a flagship. DeepSeek is retiring V4-Flash and V4-Flash-Vision-Exp now, and will route deepseek-v4-pro to V4.1-Flash at Flash rates from 04:00 UTC on September 14, 2026, until a future V4.1-Pro. If you still pin Pro by name, you get Flash billing after that cutover.
What changed in the architecture
The model card (opens in new tab) describes a 40-layer Transformer split as a 20-layer causal encoder plus a 20-layer decoder. Decoder KV is projected from the encoder’s last hidden states, not rebuilt per decoder layer. Prefill is cheaper (8B active). Decode is 16B. MoE layout: 1 shared expert and 384 routed experts per layer, 6 routed experts active per token.
Compressed Sparse Attention 2 plus FP4 main KV caching (E2M1, one E4M3 scale per 16 channels) is how they get the cache number people will repeat: 890 bytes of global KV per token. DeepSeek says that is about 1/4 of V4-Flash and about 437× smaller than V1. The launch post also says 1/4 the HBM and 1/8 the SSD versus the previous generation. Cache-hit tokens are a large share of agent bills; shrinking the cache is the point.
Other card details worth keeping as DeepSeek-reported, not independently rerun: Engram conditional memory (196B parameters, sparse lookup), DSpark speculative decoding, a from-scratch DeepSeek-ViT, 45T multimodal pretrain tokens, context extended to 1M at 34T tokens. Instruct models expose a reasoning effort integer from 1 to 100. All instruct scores below use effort 100.
A GitHub repo named DeepSeek-V4.1-Flash was not public when checked September 10, 2026. The weights and the technical report PDF (opens in new tab) are on Hugging Face.
DeepSeek’s own table is mixed
DeepSeek-reported instruct scores, max effort, temperature=1.0, top_p=0.95. Agentic coding benches used DeepSeek Harness Minimal with a 1M window unless noted. These are vendor numbers.
| Benchmark | Opus 5.0 | GPT-5.6 Sol | V4-Pro | V4.1-Flash |
|---|---|---|---|---|
| GPQA Diamond | 93.4 | 94.1 | 92.4 | 90.9 |
| HLE | 56.3 | 44.5 | 42.7† | 36.8 (39.1†) |
| Terminal-Bench 2.1 | 89.1 | 88.8 | 87.9 | 90.6 |
| Terminal-Bench 3.0 | 43.3 | 34.4 | 11.8 | 30.0 |
| Terminal-Bench 4.0 | 51.8 | 39.9 | 12.4 | 31.2 |
| DeepSWE v1.1 | 74.0 | 73.0 | 62.7 | 74.2 |
| ProgramBench | 37.0 | 23.0 | 15.5 | 20.3 |
| NL2Repo-Bench | 75.3 | 56.8 | 61.5 | 64.0 |
| CyberGym | - | 84.5 | 83.3 | 88.1 |
| SEC-Bench Pro | - | 74.3 | 56.4 | 62.8 |
| ExploitGym | 22.1 | 33.7 | 5.4 | 15.3 |
| AutomationBench | 50.3 | 45.8 | 43.2 | 54.8 |
| Agent’s Last Exam | 28.6 | 26.7 | 25.7 | 31.8 |
| HLE w/ tools | 63.6 | - | 60.0 | 63.9 |
† Text-only HLE subset.
V4.1-Flash is not “ahead of flagship models” as a blanket claim. It leads DeepSeek’s table on Terminal-Bench 2.1, DeepSWE, CyberGym, AutomationBench, and Agent’s Last Exam. It trails Opus 5.0 badly on Terminal-Bench 3.0 and 4.0, ProgramBench, NL2Repo, and HLE. GPT-5.6 Sol still owns GPQA Diamond, SEC-Bench Pro, ExploitGym, and ZeroBench-main in that same card.
Scaffolding moves the DeepSWE number: mini-SWE 74.2, DSH Minimal 72.6, Claude Code 69.8, Codex 65.6. If you quote 74.2, you are quoting mini-SWE, not a harness-invariant result.
LMArena’s text leaderboard (opens in new tab), checked September 10, 2026, lists DeepSeek V4-Pro and V4-Flash variants. It does not list deepseek-v4.1-flash. Artificial Analysis had no public 4.1-Flash model page (404). Say not listed on the public leaderboards checked, not “no benchmarks exist.”
API: deepseek-flash, 1M context, vision, cheap cache hits
Pricing (opens in new tab) is per million tokens. New rates took effect 04:00 UTC on September 10, 2026. Off-peak is half of peak. Peak windows: 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday.
deepseek-flash:
- Off-peak: cache hit $0.003, cache miss $0.15, output $0.60
- Peak: cache hit $0.006, cache miss $0.30, output $1.20
Context 1M. Max output 384K. JSON, tools, Responses API, Anthropic-format API, vision. FIM is non-thinking only. Thinking is the default; see DeepSeek’s thinking-mode guide (opens in new tab). Concurrency limit 2500 vs 500 on the still-listed Pro id.
Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp already route to V4.1-Flash at Flash prices.
Named synthetic bill, no tools, no subscription: 1M cache-miss input + 200k output.
- Flash off-peak: $0.27
- Flash peak: $0.54
- Same shape at GPT-5.6 Sol $4 / $20: $8.00
- Claude Fable 5.1 $10 / $50: $20.00
Cache hits change the story. 1M hit + 200k output off-peak is $0.123. That is why DeepSeek keeps talking about KV bytes.
OpenAI-format base URL https://api.deepseek.com. Anthropic-format https://api.deepseek.com/anthropic. DeepSeek says WorkBuddy (including CodeBuddy) and OpenCode already support the new id.
What to do with it
If you run agents on DeepSeek today, switch the id to deepseek-flash before September 14, when Pro requests become Flash anyway. Re-test Terminal-Bench 3/4 style long computer-use, not only TB 2.1. The vendor table says the cheap model wins some agent loops and still loses the newer terminal benches to Opus 5.0.
If you self-host, the card points at the inference folder for weight conversion. Recommended sampling: temperature 1.0, top_p 0.95 or 1.0, 1M context, max_tokens at least 256K. There is no Jinja chat template in this release; encoding is a Python reference plus deepseek-recipe (opens in new tab).
Independent scores are still pending. Treat the Hugging Face table as DeepSeek-reported, with harness footnotes attached.