Wait Which Model?
← Back to directory
DeepSeek

DeepSeek-V4.1-Flash

Released Sep 10, 2026 · knowledge cutoff unpublished

Status
Frontier
Location
China
Modality
Multimodal
Context window
1M
Max output
384K
Speed (max effort)
222 tok/s · 1.16s to first answer token
Price ($/MTok in / out)
$0.15 / $0.60
Cost per task (max effort)
$0.27
Open weights
Yes
Benchmarks4 of 10 reported

GPQA Diamond

90.9%

Terminal-Bench 2.1

90.6%

Humanity's Last Exam

36.8%

GDPval-AA v2

1632

MMLU-Proretired

—

SWE-bench Verified

—

SWE-bench Pro

—

AIME (latest)retired

—

LMArena Elo

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Reading is cheaper than writing by design — the encoder half activates half the parameters of the decoder and the KV cache is a quarter of V4-Flash's, so re-reading a large context each agent turn is where the savings actually land
  • Vision was trained in from the start of pre-training rather than bolted on as with V4-Flash-Vision-Exp — screenshots and charts go through the same model that does the coding, with no separate endpoint
  • Reasoning effort is a continuous 1–100 dial instead of three fixed levels, so token spend can be tuned per call rather than per mode
  • Runs on far less hardware than the parameter count suggests — the 197B Engram lookup table can sit on SSD, and Ollama, Baseten and DeepInfra served it on day one under MIT

Weaknesses

  • Thinks at enormous length on max effort — Artificial Analysis found it far more verbose than its peers, so the per-token discount arrives at the bill diluted and its per-task cost is roughly ten times V4-Flash-0731's
  • Builds plausible-looking scaffolding fast but the logic underneath breaks under scrutiny — hands-on tests caught a Rubik's-cube 'solver' that just replayed the scramble in reverse
  • Agent scores swing with the scaffold — DeepSeek's own table shows several points between its harness and Claude Code or Codex, so the headline figures need the matching harness to reproduce
  • Still answers rather than abstains when it doesn't know — Artificial Analysis measures a near-total hallucination rate on its knowledge probe, unchanged from the V4 line

DeepSeek's first Causal Encoder-Decoder model: a 552B-backbone MoE (763B on disk with the 197B Engram memory, DSpark draft module and vision encoder) activating 8B parameters during prefill and 16B during decode, trained from scratch on 45T multimodal tokens. GPQA Diamond 90.9, Terminal-Bench 2.1 90.6 and HLE 36.8% are DeepSeek's own figures at max reasoning effort (reasoning_effort=100) in its minimal-mode DeepSeek Harness; HLE is 39.1% on the text-only subset and 63.9% with tools, and the same weights score 88.0 on Terminal-Bench 2.1 under Claude Code and 84.1 under Codex per DeepSeek's cross-scaffold table. Artificial Analysis independently measures HLE 39.2% and rates it 1632 on GDPval-AA v2; it scores 40 on AA's Intelligence Index v4.3, which was rescaled since the 50 recorded for V4-Flash-0731 and no longer includes GPQA, so the two index figures are not comparable. Also reported by DeepSeek: DeepSWE v1.1 74.2, Terminal-Bench 3.0 30.0 and 4.0 31.2, Codeforces 3471, MathArena Apex 65.6, NL2Repo-Bench 64.0, CyberGym 88.1, SEC-Bench Pro 62.8, AutomationBench 54.8, Agents' Last Exam 31.8, Chartography 78.9, BabyVision 89.6 and ZeroBench 49.0 (the last three with tools). MMLU-Pro is published only for the base model (74.1, 5-shot) and is not recorded as an instruct score; SWE-bench Verified/Pro, AIME, LMArena and ARC-AGI-2 are unpublished, and Vals AI ranks it the top open-weight model on its index at high effort. Pricing follows the off-peak convention of the V4-Flash-0731 entry: $0.15 input (cache miss) / $0.60 output with cache hits at $0.003, doubling to $0.30 / $1.20 / $0.006 during peak hours (01:00-04:00 and 06:00-10:00 UTC, Monday-Friday) — the peak rate is what Artificial Analysis lists. costPerTask ($0.2652) and speed (222.2 tok/s, 1.16 s to first answer token) are AA's max-effort measurements on DeepSeek's API. Context is the 1M DeepSeek's pricing page states; max output is that page's 384K maximum, while the Hugging Face card recommends max_tokens of at least 256K. The API id is `deepseek-flash`; DeepSeek's change log retired V4 Flash and V4 Flash Vision Exp the same day and temporarily routes `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` to this model at Flash rates, which is the explicit replacement behind predecessorId. The launch post also said all `deepseek-v4-pro` traffic would route here from 2026-09-14 until V4.1-Pro ships, but the change log and pricing page later reversed that: V4 Pro service continues after 2026-09-14 with billing unchanged. As with the V4 line, no Jinja chat template ships — prompts go through DeepSeek's Python encoding reference or its Rust deepseek-recipe toolkit.

For developers

API model strings

  • deepseekdeepseek-flash

Licence

MIT License · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

Replaces DeepSeek-V4-Flash-0731

News