Recommended Articles
Tech & AI

DeepSeek Releases DeepSeek-V4.1-Flash-1017: Open Weights, Benchmark Gains, and 20-Cent Coding Tasks

By Angeled · October 19, 2026 · Tech & AI
DeepSeek Releases DeepSeek-V4.1-Flash-1017: Open Weights, Benchmark Gains, and 20-Cent Coding Tasks
DeepSeek Releases DeepSeek-V4.1-Flash-1017: Open Weights, Benchmark Gains, and 20-Cent Coding Tasks

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, a post-training update to the DeepSeek-V4.1-Flash model that launched on September 10, 2026. DeepSeek calls it a "minor update" — but open weights, benchmark gains, and aggressive pricing make it one of the year's most consequential model releases. It went live the same day on DeepSeek's website, app, and API under the unchanged API identifier deepseek-flash, at the same price. The weights were published to Hugging Face and ModelScope under the MIT license, permitting commercial use.

The release, in brief

For existing API users, the update arrived automatically: the deepseek-flash API name now serves the new checkpoint, with pricing unchanged.

The model card claims improvements across reasoning, coding, agentic, and visual benchmarks, plus reduced task execution time and shorter output lengths — the two ingredients that lower per-task cost. Context stays at 1 million tokens, with max output up to roughly 384K tokens. Architecturally, it is a mixture-of-experts design with roughly 8–16 billion parameters active per token: the full model is far larger, but only a slice fires on any token — the standard recipe for serving big models cheaply.

The benchmarks: DeepSeek's own numbers

Every figure below comes from DeepSeek's own published evaluations — none independently verified. DeepSeek compares the 1017 checkpoint against the September version and against rivals:

| Benchmark | Flash-1017 | Flash (Sept) | Claude Opus 5 | GPT-5.6 Sol | | --- | ---: | ---: | ---: | ---: | | DeepSWE v1.1 (coding) | 77.0 | 74.2 | 74.0 | 73.0 | | Terminal-Bench 4.0 (agent coding) | 42.3 | 31.2 | 51.8 | — | | Automation-Bench v1.0.6 | 60.5 | 54.8 | 50.3 | — | | Agents' Last Exam | 35.4 | 31.8 | 28.6 | — | | HLE without tools (reasoning) | 42.6 | 36.8 | 56.3 | 44.5 | | ProgramBench | 29.1 | 20.3 | — | — | | NL2Repo-Bench | 68.2 | 65.4 | — | — |

The pattern: Flash-1017 beats Opus 5 on coding and agentic benchmarks (DeepSWE, Automation-Bench, Agents' Last Exam) but trails it decisively on raw reasoning — HLE at 42.6 vs 56.3, Terminal-Bench at 42.3 vs 51.8 — and Opus 5 also leads on ProgramBench and NL2Repo-Bench. DeepSeek says the new model beats the old one on all 11 benchmarks it published, by 2.0 to 11.1 points.

The caveat is load-bearing. These scores were produced on harnesses DeepSeek chose, and independent evaluator Vals AI previously found DeepSeek's headline coding score inflated by its own harness. Treat the table above as company-reported until third-party boards reproduce it.

What the MIT license means for builders

The weights ship under the MIT license, about as permissive as open licensing gets: download, modify, and use commercially — including fine-tuning and redistribution — with minimal obligations (keep the copyright notice). Teams can test DeepSeek's hosted API while running the identical weights on their own infrastructure, with no contractual dependency on DeepSeek.

The fine print: the release totals over 500 GB (self-hosting needs serious GPUs), and the MIT license doesn't indemnify you.

Explainer: KV-cache compression and the 1M context

The companion paper is titled "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression" — the memory story is on the cover. A short explainer for non-engineers:

When a model reads a long document, it stores a key-value (KV) cache — notes about every token already read, so it isn't recomputed while generating an answer. For a 1-million-token context, that cache can grow enormous. DeepSeek compresses it aggressively (roughly 890 bytes per token), which is what makes a 1M-token context cheap enough to serve at $0.15 per million input tokens. Combined with the mixture-of-experts design activating only 8–16B parameters per token, the result is long-context inference at budget-model prices.

New Search and Fetch APIs

DeepSeek also opened a Search API and a Fetch API to all API users, built on its in-house search engine, so agents can pull live web content directly: $0.75 per 1,000 successful requests, capped at 5 QPS per account. Server-side calls via the DeepSeek Harness, Claude Code, or Codex stay free until 16:00 UTC on October 31, 2026 (November 1, Beijing time), when billing begins.

The core rate card: $0.15 per million input tokens, $0.60 per million output tokens, $0.003 per million cached-input tokens off-peak, doubling on weekdays 01:00–04:00 and 06:00–10:00 UTC. DeepSeek's own chart claims roughly $0.20 per DeepSWE task versus $2 to $10+ for rivals — vendor-produced, but the underlying rate card is public and checkable.

Why it matters for founders and investors

For founders, this is how fast AI infrastructure commoditizes: near-top coding-agent scores — company-reported, but consistent with the pricing math — at roughly one-fortieth the per-token cost of flagship alternatives, with MIT weights for commercial use. Budget a "verification tax," though: vendor scorecards are marketing until independently reproduced, and the weights swapped silently — deepseek-flash started serving the new checkpoint before documentation appeared, so consumers absorbed a behavior change with no changelog.

For investors, the signal is the price-performance trajectory, not any single score. Every few months, the cheapest credible model moves closer to the frontier. The AI-infra winners of this cycle will be the ones whose cost curves survive commoditization — distribution, proprietary data loops, and workflow lock-in — rather than raw model quality alone.

Angeled.com connects founders building in this space with investors.

Read Next