DeepSeek-V4.1-Flash-1017: a "minor update" that moves the cost frontier

TL;DR. DeepSeek released V4.1-Flash-1017 on 2026-10-17 as open weights under MIT, labelled a minor update to V4.1 Flash (09-10) (552B total, 8B active reading input and 16B active writing output, a compressed global KV cache of about 890 bytes per token). Same architecture, new post-training: reinforcement learning with self-feedback on end-to-end project building. The headline is not one benchmark. It is that the model scores higher and finishes tasks faster with shorter outputs, which DeepSeek calls a Pareto improvement in performance and cost. It also opened its in-house Search and Fetch APIs at $0.75 per 1,000 calls.

Same weights budget, fewer tokens per finished task

The architecture did not change. Post-training cut output length, which is where agent cost lives.

flowchart LR
  B["V4.1 Flash<br/><small>552B, 8B/16B active</small>"] --> R["Self-feedback RL<br/><small>whole-project tasks</small>"]
  R --> M["Flash-1017<br/><small>MIT open weights</small>"]
  M --> Q["Higher scores<br/><small>TB4 31.2 to 42.3</small>"]
  M --> T["Shorter outputs<br/><small>faster task completion</small>"]
  T --> C["Lower cost/task<br/><small>Pareto shift</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class B input
  class R loop
  class M core
  class Q,T,C exit

Blue is the base model, amber is the training change, purple is the release, green is the result.

Headline numbers (DeepSeek's table)

Benchmark	1017	V4.1 Flash	GPT-5.6 Sol	Opus 5
Terminal-Bench 4.0	42.3	31.2	39.9	51.8
DeepSWE v1.1	77.0	74.2	73.0	74.0
ProgramBench	29.1	20.3	23.0	37.0
Automation-Bench	60.5	54.8	45.8	50.3
Apex Bench v1.1	72.1	64.6	54.3	65.8
HLE (no tools)	42.6	36.8	44.5	56.3

Terminal-Bench uses the Claude Code harness; DeepSWE uses Mini-SWE-V2. Context 1M tokens, recommended max_tokens ≥256K. No Jinja chat template: a Python reference encoder plus a Rust deepseek-recipe library handle prompt formatting.

Why it matters

Token optimization as a release axis. Thom Wolf's read on X: shorter outputs are "one of the practical ways the cost curve collapses, often done as an afterthought." DeepSeek plots DeepSWE pass@1 against cost per task, not against parameters.
Harness-tuned. Optimized for DeepSeek Harness v0.2.1+, Tencent CodeBuddy and WorkBuddy. Benchmarks are reported per harness, matching the wiki's thread that the harness moves cost as much as the model (HarnessRouter, 10-17).
Interconnects (10-17) places V4.1 Flash near Meta's Muse Spark Open on capability. 1017 lands two days after Muse Spark Open, widening the gap Meta had just narrowed.
Search and Fetch at $0.75 per 1,000 calls, free inside DeepSeek Harness, Claude Code and Codex until 2026-10-31. A lab bundling its own retrieval with cheap tokens.

Gaps

All numbers are DeepSeek's own. No output-length or cost-per-task figure is given in text, only a chart. The tech report title ("Pushing the Limits of KV Cache Compression") is the 09-10 report; 1017 adds no architecture detail.

Source: Hugging Face model card · DeepSeek news · The Decoder