DeepSeek V4.1-Flash-1017

DeepSeek is calling it a minor update. The numbers tell a louder story. On Friday the Hangzhou lab published DeepSeek-V4.1-Flash-1017 — a refreshed checkpoint of its lightweight flagship — with weights on Hugging Face and ModelScope under an MIT license, and its website, app, and API all switched over to the new model the same day.

What was released

The model card describes 1017 as “a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks.” Under the hood it’s a 552-billion-parameter sparse mixture-of-experts model — the first built on DeepSeek’s Causal Encoder-Decoder (CED) architecture — activating just 8B parameters per input token and 16B per output token, with a 1,048,576-token context window. Image understanding is native to the architecture, trained jointly from the start of pre-training rather than bolted on afterward.

The headline benchmark scores, all DeepSeek’s own: 42.3 on Terminal-Bench 4.0 (up from 31.2) and 77.0 on DeepSWE v1.1 (up from 74.2) — the latter ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on that specific coding benchmark, per the published comparisons. DeepSeek also reports shorter outputs and faster task completion, calling it a Pareto improvement in performance and cost. Independent evaluator Vals AI ranks it the number one open-weight model on its index, though it also found DeepSeek’s headline coding score was inflated by its own harness — a reminder to read vendor numbers carefully.

The price war continues

Pricing on DeepSeek’s own API is unchanged: $0.15 per million input tokens and $0.60 per million output tokens, under the same deepseek-flash alias — existing callers get the upgrade silently. Alongside the weights, DeepSeek opened its Search and Fetch APIs to everyone at $0.75 per 1,000 requests (capped at 5 QPS per account), backed by an in-house search engine. A companion deepseek-recipe library (Rust with Python bindings) handles prompt conversion and output parsing, and the companion paper — “DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression” — puts the memory-efficiency story front and center.

Why it matters

The announcement drew 3,600 likes in under an hour on X, with one reply summing up the mood: “the whale is back.” Timing is everything: V4.1 Flash originally launched September 10, and just a day before this refresh, the coding agent OpenCode made it free to use. DeepSeek is now shipping improved weights roughly monthly while holding prices flat — compressing the cost curve for coding agents exactly as U.S. labs push their own open-weight releases.

Whether the 1017 checkpoint holds its benchmark leads under independent testing is the open question. The weights are out there for anyone to check.

Sources: Smart Chunks (Oct 17, 2026); AI Weekly (Oct 17, 2026); AI News Daily (Oct 17, 2026). Benchmark figures are DeepSeek’s own and await independent verification.