DeepSeek Releases V4.1-Flash-1017: MIT-Licensed Weights, In-House Search API

What happened

On October 17, DeepSeek released V4.1-Flash-1017, an updated snapshot of the V4.1-Flash launched in September, following the date-suffix convention of V4-Flash-0731 and V4-Pro-0813. The model went live on the website, app and API, served under the existing name deepseek-flash; the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work and are served by the new snapshot. Model API pricing is unchanged: $0.30 per million input tokens and $1.20 per million output tokens at peak, half that off-peak. The company calls it 'a Pareto improvement in performance and cost.' The model card describes the release as a minor update to V4.1-Flash, with gains across reasoning, coding, agentic and visual benchmarks, plus shorter task execution time and output length.

All published benchmark figures are DeepSeek-reported and have not been independently reproduced. DeepSWE v1.1 scored 77.0 (previous generation 74.2; DeepSeek's table lists Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0); Terminal-Bench 4.0 scored 42.3 (previous generation 31.2; Opus 5 at 51.8 and GPT-5.6 Sol at 39.9), which DeepSeek says it ran with the Claude Code harness; HLE without tools scored 42.6 (Opus 5 at 56.3). The model card lists 11 benchmarks, each 2.0–11.1 points above the previous generation.

DeepSeek also opened its in-house Search and Fetch API the same day. The official docs show: Search (POST /v1/tools/search) returns 1–50 structured web results, supporting site:, lang:, after: and before: operators, with a 0–3 rerank score and a 0–4 authority tier per result; Fetch (POST /v1/tools/fetch) converts a single webpage into Markdown. Both endpoints are priced identically: $0.75 per 1,000 successful requests, billed only on HTTP 200 with results produced — failures, empty results and timeouts are free; each endpoint is rate-limited to 5 QPS per account. Server-side Search/Fetch calls inside DeepSeek Harness, Claude Code and Codex are currently free, with per-request billing starting 00:00 Beijing time on November 1, 2026. DeepSeek says its in-house search engine serves hundreds of millions of requests daily.

Key facts

1. Release: October 17, 2026 — an updated snapshot of the September V4.1-Flash, live on the website, app and API
2. Weights license: The Hugging Face model card lists both code and weights under the MIT license (some outlets' spec tables said 'not disclosed'; the model card is authoritative)
3. Benchmarks (DeepSeek-reported): 77.0 on DeepSWE v1.1, 42.3 on Terminal-Bench 4.0, 42.6 on HLE without tools; 11 benchmarks up 2.0–11.1 points over the previous generation, not independently reproduced
4. Architecture: 552B backbone plus about 196B Engram conditional-memory tables; 8B active per token in prefill, 16B in decode; 1M-token context; FP4 KV cache at 890 bytes per token
5. Search API: Search (1–50 structured results) and Fetch (webpage to Markdown) priced identically: $0.75 per 1,000 successful requests; 5 QPS per endpoint per account
6. Free window: Server-side Search/Fetch calls inside DeepSeek Harness, Claude Code and Codex are currently free; per-request billing starts 00:00 Beijing time, November 1, 2026

Background

The V4 lineage: the V4 family preview on April 24, 2026; V4-Flash-0731 on July 31 (284B total / 13B active per token); V4-Pro-0813 on August 12–13 (1.6T total / 49B active); V4.1-Flash on September 10 (552B backbone plus about 196B Engram conditional-memory tables; 8B active per token in prefill, 16B in decode; 1M-token context; FP4 KV cache at 890 bytes per token). The accompanying technical report is titled 'DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression.' The entire V4 series is MIT-licensed on Hugging Face, and 1017 continues that practice.

The new snapshot is optimized for agent coding tools including DeepSeek Harness, Tencent's CodeBuddy and WorkBuddy, with DeepSeek recommending Harness v0.2.1-rc.1 or later. Harness is the agent framework DeepSeek open-sourced under the MIT license on August 13, reaching about 197,000 GitHub stars within eight days.

DeepSWE v1.1 is maintained by Datacurve: 113 original long-horizon software-engineering tasks across 91 repos and 5 languages, with hand-written verifiers to avoid training-data contamination — it measures whether an agent can ship a working patch. Terminal-Bench 4.0 is maintained by the Terminal-Bench team: 66 human-verified containerized terminal tasks where an agent operates in a real terminal and passes only when hidden verification scripts confirm the machine's end state.

The MIT license already produced a fork effect in September: hours after V4.1-Flash launched, abliterated/uncensored forks appeared on Hugging Face; third-party tracker shattered.io counts 31 builds with 83,900 cumulative downloads in that fork ecosystem.

On the regulatory front, THE DECODER reports the Chinese government is considering tightening or restricting open-weight model releases, reportedly using a top-10-most-capable criterion to draw the line; aiweekly notes DeepSeek frames 1017 as a 'tuned checkpoint,' sidestepping frontier-model regulatory scrutiny in multiple jurisdictions.