DeepSeek posted V4.1-Flash-1017 to Hugging Face under an MIT license on October 17, 2026, with the model card claiming 77.0 on DeepSWE v1.1, ahead of the previous Flash at 74.2, Claude Opus 5 at 74.0, and GPT-5.6 Sol at 73.0. The checkpoint reaches callers through the same deepseek-flash API identifier they were already using yesterday. No version bump, no changelog gate, no opt-out, per Data Studios and an AI Weekly editor's note.

That's the structural story. The silent swap is becoming the dominant release shape for budget-tier Chinese labs, and it collapses the usual migration ritual (pin a version, run a shadow eval, cut over) into a benchmark announcement and a prayer. It also means hosted callers inherit the gains automatically at the unchanged $0.15 per million input tokens, $0.60 per million output, and $0.003 per million on cache hits.

The gains aren't a single-benchmark fluke. MarkTechPost reports all 11 benchmarks DeepSeek published improved, by anywhere from 2.0 to 11.1 points. Terminal-Bench 4.0 climbs from 31.2 to 42.3. ProgramBench moves from 20.3 to 29.1. Automation-Bench v1.0.6 goes 54.8 to 60.5. Apex Bench v1.1 lifts 64.6 to 72.1. On HLE without tools, Flash-1017 posts 42.6 against Opus 5's 56.3, reasoning-heavy terrain where frontier closed models still hold the ceiling.

Under the hood, the companion paper Pushing the Limits of KV Cache Compression describes an asymmetric 20-layer causal encoder paired with a 20-layer decoder, an FP4 global KV weighing in at roughly 890 bytes per token (about a quarter of V4-Flash's KV footprint), and an SWA bounded replay cache at roughly one-eighth. The context window stays at 1M tokens with a 256K minimum output capacity and 384K available on the API.

The other half of the release is the Search and Fetch API. AI Weekly reports pricing at $0.75 per 1,000 successful requests, with a 5 QPS per-account cap, and server-side calls through DeepSeek Harness, Claude Code, or Codex free until November 1, 2026. Data Studios logs the billing cutover more precisely: October 31, 2026 at 16:00 UTC.

That leaves roughly a two-week window to measure retries, cache-hit rate, and completed-workflow cost on grounded calls before the meter starts. The last comparable moment was the 2023 OpenAI tool-use pricing debut, when early adopters who instrumented retry behavior arrived at November with usable unit economics while everyone else arrived with surprise invoices. The timing isn't accidental.

Sources

https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
https://arxiv.org/pdf/2609.19969
https://www.marktechpost.com/2026/10/17/deepseek-just-released-v4-1-flash-1017-scores-77-0-on-deepswe-opens-search-api/
https://aiweekly.co/alerts/deepseek-opens-v41-flash-1017-weights-and-075-search-api
https://www.datastudios.org/post/deepseek-v4-1-flash-1017-open-weights-coding-benchmarks-api-pricing