DeepSeek-V4.1-Flash-1017
Released Oct 17, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- China
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1M
- Max output
- 384K
- Speed (max effort)
- 230 tok/s · 1.08s to first answer token
- Price ($/MTok in / out)
- $0.15 / $0.60
- Cost per task (max effort)
- $0.26
- Open weights
- Yes
Strengths & weaknesses
Strengths
- Agentic task-completion and tool-use loops reportedly hold together longer than the September checkpoint, per DeepSeek's own evaluation harness
- Shorter, less sprawling answers than V4.1-Flash on the same tasks, per DeepSeek's own release notes
- Same MIT weights land on Hugging Face and third-party hosts (DeepInfra, Baseten) the day of the announcement, continuing V4.1-Flash's day-one multi-host availability
Weaknesses
- Shipped as a same-day checkpoint swap under the existing deepseek-flash API alias rather than a standalone announcement — every reported gain so far is DeepSeek's own harness output relayed by trackers, with no independent eval yet
- Still ships without a Jinja chat template, inheriting V4.1-Flash's reliance on DeepSeek's own Python/Rust prompt-encoding tooling
A same-architecture dated refresh of DeepSeek-V4.1-Flash (2026-09-10), released 2026-10-17 under the same deepseek-flash API alias, MIT license and 1M-token context; DeepSeek's own figures (relayed by MindStudio, OfficeChai, Kingy AI, MarkTechPost and aiweekly.co, none independently verified) report Humanity's Last Exam at 42.6% without tools (up from 36.8%), Terminal-Bench 4.0 at 42.3 (not comparable to this site's Terminal-Bench 2.1 field, left null) and DeepSWE v1.1 at 77.0; GPQA Diamond, SWE-bench, LMArena and ARC-AGI-2 were not reported for this checkpoint specifically and are left null pending a dedicated eval. contextWindow, maxOutput, pricing and license are carried over from the shared V4.1-Flash deployment since every source describes this as an in-place checkpoint swap on the same endpoint, not a new context/pricing tier; speed (230.4 t/s, 1.08s on DeepSeek's API) and cost per Intelligence Index task ($0.26) are Artificial Analysis' measurements of the 1017 checkpoint at max effort; Artificial Analysis lists DeepSeek's peak-hour rate of $0.30/$1.20 per MTok, while the pricing above is the off-peak rate.
Benchmarks2 of 10 reported
Humanity's Last Exam
ARC-AGI-2
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
LMArena Elo
GDPval-AA v2
For developers
API model strings
- deepseek
deepseek-flash
Licence
MIT License · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
Replaces DeepSeek-V4.1-Flash