DeepSeek V4.1-Flash-1017 Hits 77% on Coding Agents as Self-Feedback RL Cuts Task Cost

DeepSeek released an updated model checkpoint on October 17, 2026, pushing its V4.1-Flash model to 77.0% Pass@1 on the DeepSWE v1.1 coding-agent benchmark and simultaneously opening a self-built web search API to all developers. The update, labeled V4.1-Flash-1017, ships on the same deepseek-flash API identifier as its predecessor, requires no code changes from existing integrations, and arrives with MIT-licensed weights on Hugging Face and ModelScope within hours of launch.

The improvement is not from a new architecture — the underlying Causal Encoder-Decoder (CED) design introduced on September 10 is unchanged. What changed is the post-training: an extended round of reinforcement learning using execution-based self-feedback, focused on long-horizon construction tasks. The result is a model that scores higher on agentic benchmarks and generates shorter outputs, which translates directly to lower token costs per completed task.

What Self-Feedback Reinforcement Learning Actually Did

DeepSeek's technical report for V4.1-Flash, published on arXiv as "Pushing the Limits of KV Cache Compression," describes its post-training pipeline as supervised fine-tuning followed by reinforcement learning followed by on-policy distillation. The 1017 checkpoint extends the RL phase with a "self-feedback" mechanism: instead of relying on human preference labels, the model generates candidate solutions, executes them in a sandboxed environment, and uses pass/fail signals from test execution as the reward signal.

This setup is well-suited to coding and construction tasks because correctness can be automatically verified — a game either runs or it doesn't; a 3D scene either renders or it crashes. The model is penalized for verbose outputs that pass tests no better than concise ones, which is how the update produces both shorter completions and improved pass rates simultaneously. DeepSeek described it as a "Pareto improvement": both the performance curve and the cost curve moved in the developer's favor without a tradeoff between them.

The demo content DeepSeek released alongside the update illustrated the focus areas. A Palace Museum 3D walkthrough, a browser-based racing game with selectable vehicles and a lap-completion interface, a Portal 2 recreation in Godot, and an animated historical city timeline from prehistory to the modern era all appear to have been generated from zero with minimal instruction — what the source material calls "wish-driven programming," essentially the ability to describe a complex project and receive a runnable prototype. These demos are marketing artifacts rather than peer-reviewed evaluations, but they consistently test the same domain: multi-step interactive construction with a verifiable end state.

Read more: DeepSeek V4.1 Flash CED architecture explained: KV cache and agent benchmarks

Benchmarks: What the Numbers Mean and Where to Apply Caution

The V4.1-Flash-1017 model card on Hugging Face lists seven benchmark comparisons against the September V4.1-Flash baseline:

• DeepSWE v1.1: 77.0 vs 74.2 (a 3.8-point gain)

• Terminal-Bench 4.0: 42.3 vs 31.2 (a 35.6% relative gain)

• ProgramBench: 29.1 vs 20.3 (a 43.3% relative gain)

• Automation-Bench v1.0.6: 60.5 vs 54.8

• Apex Bench v1.1: 72.1 vs 64.6

• Agents' Last Exam: 35.4 vs 31.8

• NL2Repo-Bench: 68.2 vs 65.4

These are company-reported figures using what DeepSeek calls the "mini_swe_v2" harness — an internal test protocol for running DeepSWE tasks. This is not the same harness that Datacurve, the independent operator of the DeepSWE v1.1 leaderboard, uses. The absolute numbers differ between harnesses; comparisons within the same harness are valid, but cross-system comparisons require care.

With that caveat stated, the competitive picture is directionally consistent. Datacurve's publicly published leaderboard shows the top cluster of frontier models — GPT-6 Astra, Gemini-3.8-Flash, and Claude Opus 5 — each scoring approximately 74% on DeepSWE v1.1 in their respective high-effort configurations. DeepSeek claims 77.0% for the 1017 update in its max configuration, a 3-point lead over that cluster using its own internal harness. The 1017 checkpoint has not yet appeared on Datacurve's independently run leaderboard as of this writing.

The cost dimension is where the claim becomes harder to dispute. Datacurve's leaderboard shows Claude Opus 5 at $11.84 per task and GPT-6 Astra at $4.43 per task on equivalent evaluations. DeepSeek's cost chart for the 1017 update places its max configuration at approximately ¥1–2 per task ($0.14–$0.28 at current exchange rates), down from ¥3–4 per task for the September Flash baseline. If the benchmark positions hold in independent testing, the cost-performance gap between V4.1-Flash-1017 and closed-source competitors is not a rounding difference — it is roughly an order of magnitude.

Shorter model outputs are the primary mechanism for the cost reduction. The RL training penalized verbosity, so the model learned to complete tasks with less generated text. Because token costs are linear in output length, and because agentic tasks typically involve many model calls, this compounds: fewer tokens per call, across many calls, reduces both cost and wall-clock latency.

A Self-Built Search Index Joins the API

Simultaneous with the model update, DeepSeek opened two new REST endpoints: Search (POST /v1/tools/search) and Fetch (POST /v1/tools/fetch). The Search API documentation describes an endpoint that queries DeepSeek's own web index and returns structured results with relevance scores, page summaries, and metadata. Fetch takes a URL and returns its content as Markdown. Together, they give coding agents and retrieval-augmented applications a first-party, DeepSeek-native retrieval layer without requiring a third-party search provider.

Pricing is $0.75 per 1,000 successful requests on each endpoint. Failed calls, empty results, and timeouts are not charged. The API documentation describes anti-manipulation filtering for SEO gaming, AI-generated content farms, and GEO manipulation, though no independent audit of these filters has been published.

One timing detail developers should note: server-side Search and Fetch tool calls originating from DeepSeek Harness, Claude Code, and Codex are currently billed at zero. That free window closes at 00:00 Beijing time on November 1, 2026 (16:00 UTC on October 31), after which standard per-call rates apply.

The web index scale claimed by DeepSeek in its marketing materials is described as "hundreds of billions of pages" capable of processing hundreds of millions of searches per day, with multilingual support. These figures have not been independently verified, and the API documentation itself does not include index-size disclosures or source-corpus descriptions. Developers evaluating the search endpoint for production use should run their own quality assessments on their target query domains before the free period ends.

The Open-Weights Ecosystem and What Inherits the Update

MIT-licensed weights for V4.1-Flash-1017 are available on Hugging Face under the deepseek-ai organization and on ModelScope. The model uses FP8, BF16, I8, and FP4 tensor formats across its 763B total parameters, consistent with the CED architecture's quantization-aware training approach.

Within hours of the October 17 release, 13 community quantizations appeared in the Hugging Face model tree, covering GGUF-format variants for llama.cpp, Ollama, LM Studio, and Jan. This rapid quantization response indicates active developer interest in running the model locally — notable for a 763B parameter MoE that requires substantial infrastructure at full precision, but more tractable at 4-bit quantization through sparse activation patterns that keep only 8B–16B parameters active per forward pass.

The API routing remains unchanged: deepseek-v4-flash and deepseek-v4-flash-vision-exp continue routing to the new checkpoint at Flash pricing. The update is deployed to DeepSeek's web interface and mobile app as well as the API.

Where the Technology Still Falls Short

Two limitations from the original V4.1-Flash technical report remain relevant to the 1017 update. First, the architecture paper explicitly notes that "a distinct overall performance gap remains" on science-oriented agentic tasks compared with larger closed-source systems. The RL improvements documented in the 1017 update focused on construction tasks (games, 3D, coding projects with verifiable end states), not scientific reasoning or research-oriented agent work. Second, the SWA Bounded Replay mechanism used to compress the KV cache is an approximation; reconstructed decoder SWA keys and values are not mathematically equivalent to a full forward pass. The authors describe the quality impact as negligible, but it remains an unverified approximation.

More broadly, the competitive benchmark comparison between DeepSeek's internal mini_swe_v2 harness results and Datacurve's independently run leaderboard is unresolved. The 77.0% figure is company-reported; independent reproduction will determine whether the 3-point lead over the 74% cluster holds. Developers making model selection decisions for high-value agentic pipelines should wait for independent evaluations before treating the performance claim as settled fact.

Read more: How harness design affects AI coding agent benchmark scores — Harness-of-Harness explained

The Larger Implication: RL Post-Training as the Competitive Iteration Layer

The more consequential observation from the 1017 release is architectural in a different sense than the CED design. The underlying model — 763B parameters, CED, FP4 KV quantization, 1M token context — was fixed in September. All the benchmark gains in the October update come from extending the reinforcement learning post-training phase. DeepSeek did not build a new model to improve performance; it ran more training on the existing one, targeted at specific task domains.

This is the direction the competitive frontier is moving. Architecture is increasingly a medium-term constant; the RL post-training cycle, focused on verifiable-outcome domains where self-feedback can substitute for human labels, is where near-term model improvement is being concentrated. The cost of that RL cycle is far lower than a full pretraining run, and it can be repeated on a checkpoint-by-checkpoint basis as DeepSeek identifies performance gaps.

The November 1 billing date for the search API is the next near-term milestone developers need to track. Before that date, organizations running retrieval-augmented coding agents through Claude Code or Codex against DeepSeek's search backend are receiving a subsidized service. Whether the API's retrieval quality justifies the cost at scale — relative to alternatives — is a question the free window was designed to let developers answer for themselves.