Creation & workflows
DeepSeek V4.1-Flash-1017: a Cheaper Coding Model, Now With Its Own Web Layer
On October 17, DeepSeek published the DeepSeek-V4.1-Flash-1017 checkpoint, an updated post-training run of its September V4.1-Flash, now served under the existing deepseek-flash model name. Alongside the model, DeepSeek opened its Search and Fetch APIs — powered by an in-house search engine — to every developer. For agent builders, the release is two things at once: a Pareto move on coding-agent cost, and a first-party web layer that competes with standalone search-API vendors.
Muse-assisted drafting; facts and all eight translations reviewed by the owner. This is not an independent model review. This article was drafted with AI assistance and reviewed by an AI fact-checker under owner authorization. The owner has not personally reviewed it; no independent human verification is claimed. Benchmark figures are DeepSeek’s own, with no independent reproduction observed; API and pricing facts come from DeepSeek’s public model card and API documentation.
A checkpoint, not a new model
DeepSeek-V4.1-Flash-1017 is an updated post-training checkpoint of the V4.1-Flash that launched in September — not a new architecture. It is live on DeepSeek’s website, app and API under the existing deepseek-flash model name; the legacy identifiers deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired but still accepted, served by the new checkpoint and billed at the Flash price. DeepSeek frames it as a Pareto improvement in performance and cost: stronger scores, faster task completion and shorter outputs. The 1017 checkpoint is tuned for DeepSeek Harness — v0.2.1-rc.1 or later is recommended — as well as Tencent CodeBuddy and WorkBuddy, with reinforcement learning with self-feedback for end-to-end project work.
The Hugging Face card for V4.1-Flash states that the repository and the model weights are licensed under the MIT License. Third-party coverage marks the 1017 checkpoint weights’ license as not disclosed, so whether the MIT statement covers the 1017 weights specifically was not separately confirmed on the pages observed — a detail to verify before treating the checkpoint as MIT-licensed in a compliance-sensitive deployment.
The memory story is the headline
The companion technical report is titled “DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression”, and the title is the thesis. The model is a multimodal mixture-of-experts with a 552B-parameter backbone built as a 40-layer Causal Encoder-Decoder Transformer — a 20-layer causal encoder followed by a 20-layer decoder — whose decoder global KV cache is projected from the final encoder hidden states. That design activates only 8B parameters per token during prefill and 16B during decode. Compressed Sparse Attention 2 and FP4 (E2M1) main KV caching bring the global KV footprint to 890 bytes per token, roughly a quarter of V4-Flash; SWA Bounded Replay shrinks the persistent KV footprint to roughly an eighth. Alongside sit 196B Engram conditional-memory parameters, a 45T-token pre-training corpus, a 1M-token context window and up to 384K output tokens.
Do the arithmetic: at 890 bytes per token, a full 1M-token context holds roughly 0.9 GB of persistent KV cache — a line item smaller than the model’s own weight footprint. For input-heavy agentic workloads, the binding constraint moves from VRAM capacity to orchestration and tool-call design. That is the economic argument for the whole release: long context stops being a memory luxury and becomes an architectural choice you can price.
Benchmarks, footnotes attached
DeepSeek’s own table for the 1017 checkpoint reports DeepSWE v1.1 at 77.0, up from 74.2 and ahead of Claude Opus 5’s 74.0 and GPT-5.6 Sol’s 73.0; Terminal-Bench 4.0 at 42.3, up from 31.2 but still behind Opus 5’s 51.8; Automation-Bench v1.0.6 at 60.5; Agents’ Last Exam at 35.4; Apex Bench v1.1 at 72.1; ProgramBench at 29.1; NL2Repo-Bench at 68.2; and HLE without tools at 42.6 versus 56.3 for Opus 5. Across the 11 benchmarks DeepSeek reported, the checkpoint beats the September Flash by 2.0 to 11.1 points.
Every one of these figures is DeepSeek’s own, run in its internal framework, and no independent reproduction has been published; the competitor scores are DeepSeek’s reported comparisons, not independently verified ones. That does not make the table useless — the direction and the deltas are real information — but it does mean the numbers are directional until someone else runs the harness. The durable, contractual facts in this release are the API surface and the price card, not the leaderboard.
A first-party web layer for agents
The infrastructure half of the release may matter more than the model. DeepSeek opened its Search and Fetch APIs to every developer: POST /v1/tools/search and POST /v1/tools/fetch on api.deepseek.com. Billing is per successful request — $0.75 per 1,000 requests for each endpoint — and a request counts only when it returns HTTP 200 with a result; failed requests, empty search results and timeouts are free. Each API is limited to 5 QPS per account. Server-side calls made through DeepSeek Harness, Claude Code or Codex stay free until per-request billing begins at 00:00 Beijing time on November 1, 2026.
The Search API returns structured results from DeepSeek’s in-house engine: 1 to 50 results per query, operators such as site:, lang:, after: and before:, a 0-to-3 rerank score and a 0-to-4 authority tier on each result. The Fetch API retrieves a single page and returns its title plus content as Markdown. Put behind the same API key and balance as the model, this makes the agent stack — model plus retrieval — a single billable surface, and it puts DeepSeek in direct competition with standalone search-API vendors for agent workloads.
What to do with it
Model pricing is unchanged: peak $0.30 per 1M cache-miss input tokens ($0.15 off-peak), $1.20 per 1M output tokens ($0.60 off-peak), with off-peak rates at half of peak. Nothing about the release asks you to migrate — the model name did not change, the price did not change, and the checkpoint arrives as an in-place improvement. The one calendar date that matters is November 1: free server-side search calls end, and the web layer becomes metered.
For teams running coding agents at scale, this is a procurement event as much as a model event. The right evaluation is measured cost per completed task inside your own harness, not the vendor’s leaderboard table — and, while you are at it, a real comparison of the $0.75-per-thousand search tier against whatever browsing provider your agents use today.
Sources and further reading
Source records were reviewed by the owner; no independent factual or model review is claimed.
- DeepSeek-V4.1-Flash — model card (DeepSeek, Hugging Face)
DeepSeek (Hugging Face model card)
Recorded publication date ·
Recorded verification time ·
- DeepSeek API docs — Models & Pricing
DeepSeek (API docs)
Source publication date not verified
Recorded verification time ·
- DeepSeek API docs — Search & Fetch guide
DeepSeek (API docs)
Source publication date not verified
Recorded verification time ·
- MarkTechPost — independent coverage
MarkTechPost
Recorded publication date ·
Recorded verification time ·