On October 17, DeepSeek shipped a checkpoint update to V4.1-Flash, the model it first released on September 10. The new weights landed on HuggingFace and ModelScope under the name DeepSeek-V4.1-Flash-1017, still under the MIT license and still free to download. The API side is even simpler: the service name remains deepseek-flash, and the old model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted although retired, billed at Flash rates. In other words, you change zero lines of code and your requests are already running on the new weights.
This is a textbook case of swapping the engine without touching the chassis: architecture unchanged, API name unchanged, config files identical to the September 10 build. Only the weights themselves changed. The official changelog spells it out in the October 17 entry at api-docs.deepseek.com/news/. Every number below carries its source, and anything the vendor self-reports is labeled as vendor-reported.
1. Timeline: where this update sits
Arrange the last forty days on one line. September 10: V4.1-Flash debuts and is open-sourced. October 15: the DeepSeek API suffers an outage of nearly five hours; the company later apologizes and fully refunds user spending from October 8 through 14 as account credit (per Zhidx). October 17: the 1017 checkpoint goes live with weights published in parallel.
A narrative online says the model was rushed out to offset outage backlash; that is self-media storytelling, not evidence. The confirmed facts: the outage on October 15, the official apology and refund, and the 1017 build two days later — no causal link has been confirmed, and this article will not speculate.
What actually matters to users: this is a checkpoint-level update, not a major version bump. Deployment scripts, inference configs, and quantization setups from the September 10 build should work as-is, so the upgrade cost is mostly re-downloading weights and running one regression pass.
2. Parameters: 763B total versus 8B/16B activated, keep the two apart
The headline conclusion first. Total weights are 763B (measured from the HF checkpoint), of which the backbone is 552B, the Engram conditional memory module is 196B, and the vision encoder, MLP projections, and DSpark draft model add up to roughly 15B. Activated parameters are a separate figure: 8B during prefill and 16B during decode.
Media accounts keep colliding over these numbers. Some report only 552B, others only 763B; both are incomplete. The 763B figure is the size of everything you download, which governs disk and memory planning. The 8B/16B figure is the compute actually activated per forward pass, which governs inference speed and your compute bill. You need both, and dropping either one misleads deployment decisions.
One more note on Engram: it is a standalone 196B block accessed on demand like a compressed knowledge store, not part of the usual per-layer routed activation. That is why the gap between 763B total and 8B/16B activated is so extreme: few experts wake per token, and Engram is lookup-based rather than always-on. Practical implication: do not be intimidated by 763B, and do not estimate inference cost from it either.
On disk, the checkpoint is 510 GB across 48 safetensors shards, with expert weights in FP4 and the rest in mixed precision, which is why it lands below the theoretical 763GB. Plan disk space around 510GB and leave headroom for loading.
A quick architectural sketch: a CED causal encoder-decoder with 40 layers (20 plus 20); each layer carries 1 shared expert plus 384 routed experts with 6 activated per token; the global KV cache is 890 bytes per token, about one quarter of V4-Flash and one 437th of V1; context length is 1M with a maximum output of 384K. All of this matches the September 10 build. The chassis did not move; the engine did.
The 890-byte KV cache deserves its own note: in long-context workloads the cache is the memory killer, and at 890 bytes per token the same memory fits far longer sessions or many more parallel ones — that is what makes a 1M context deployable at all. If you query large codebases, this parameter predicts your experience better than any benchmark.
3. Benchmarks: eleven gains, but Opus 5 still leads on two
The table first. All figures are vendor-reported (self-submitted by DeepSeek); no third-party board has scored the 1017 build yet:
| Benchmark | 1017 build | 9-10 build |
|---|---|---|
| Terminal-Bench 4.0 (Claude Code harness) | 42.3 | 31.2 |
| ProgramBench | 29.1 | 20.3 |
| HLE no tools | 42.6 (44.0 text-only) | 36.8 (39.1) |
| DeepSWE v1.1 (Mini-SWE-V2 harness) | 77.0 | 74.2 |
| NL2Repo-Bench | 68.2 | 65.4 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 |
| Agents' Last Exam | 35.4 | 31.8 |
| Apex Bench v1.1 | 72.1 | 64.6 |
| Chartography (with tools) | 83.0 | 78.9 |
| BabyVision (with tools) | 92.5 | 89.6 |
| ZeroBench-main (with tools) | 51.0 | 49.0 |
Three points must be made explicitly.
First, every number above is vendor-reported. Third-party trackers state plainly that no third-party board has scored the 1017 build, so the across-the-board gains currently rest on a single source. Waiting for independent runs before declaring victory costs nothing.
Second, the comparison model cannot be cherry-picked. In the vendor's own numbers, Opus 5 scores 51.8 on Terminal-Bench 4.0, above the 1017 build's 42.3, and 56.3 on HLE, above 42.6. The 1017 build does not overtake on either, trailing by 9.5 and 13.7 points respectively. The progress is real; the crown is not claimed. Both sentences belong in the same article.
Third, the gains are unevenly distributed. Terminal-Bench 4.0 jumps from 31.2 to 42.3, an 11.1-point climb and the flashiest line. ProgramBench moves from 20.3 to 29.1, up 8.8 points. Apex Bench climbs from 64.6 to 72.1. DeepSWE, already high, edges from 74.2 to 77.0. The pattern: coding and terminal-heavy benchmarks surge, while benchmarks near saturation crawl.
For orientation: Terminal-Bench 4.0 tests real-terminal task completion through the Claude Code harness — a harness is the evaluation scaffold around a model, and the same model scores differently under different ones, hence the labels in the table. HLE is hard-question answering. DeepSWE v1.1 measures software engineering under Mini-SWE-V2. NL2Repo-Bench tests operating on code repositories from natural language; Automation-Bench and Agents' Last Exam target agentic workflows. Pick the one or two rows closest to your workload and ignore the rest.
4. Price and cost: the rate card froze, the bill shrank
API pricing is unchanged: off-peak input costs 0.15 dollars per million tokens on cache miss, 0.003 dollars on cache hit, and output costs 0.60 dollars per million. Peak hours (weekdays 01:00-04:00 and 06:00-10:00 UTC) double to 0.30 and 1.20 dollars. Legacy model names bill at Flash rates, with no hidden increase.
A word on exploiting that rate card: off-peak is half price, so shift batch jobs and heavy agent runs into the valley, and cache hits are fifty times cheaper than misses — structure repeated prefixes, long system prompts, and fixed documents as cacheable prefixes. The weight update changes none of this.
The bill shrinks because of the weights. On DeepSWE v1.1 per-task cost, the best setup of the September 10 build spent about 0.50 dollars while the 1017 build spends about 0.20 (read from the vendor's chart); Zhidx converts that to roughly 3.3 yuan down to 1.3 yuan, a drop of about sixty percent (secondary source, labeled). The logic is straightforward: decode activation stays at 16B, but the new weights answer correctly more often at equal compute, so retries and rework shrink and each task consumes fewer tokens. Same task, same rate card, sixty percent off the bill. That lands harder than any benchmark table.
Also worth attention: the Search/Fetch API is now open — DeepSeek productized its in-house search engine at 0.75 dollars per thousand calls (about 5 yuan per thousand per Zhidx), temporarily free inside tools such as Claude Code and Codex. For agent builders, the model vendor now offers search and fetch natively: one less hop in the call chain, and transparent billing.
5. Wish-style coding: the new capability these weights trained in
DeepSeek gave this round of capability a name: wish-style coding. You make a wish, and it turns the project into reality, generating a runnable project from scratch when the prompt and reference material are minimal. The official demos center on game development and 3D modeling, with whole projects assembled in a few interaction rounds inside the DeepSeek Harness environment.
The training method also differs from before: these abilities come from self-feedback reinforcement learning, where the model closes the loop on its own outputs rather than piling on externally labeled data. That route corroborates the surge in terminal-oriented benchmark scores. Agentic tasks demand exactly the ability to correct oneself across rounds, not a pretty score on a single-turn exam.
On tooling: the official recommendation is DeepSeek Harness at 0.2.1-rc.1 or above; per Zhonghua Net, the vendor also optimized compatibility with Tencent CodeBuddy and WorkBuddy (secondary source, labeled). The shortest path to trying wish-style coding is upgrading the Harness to the recommended version.
A boundary here too: the official demos are game development and 3D modeling, domains where goals are visually verifiable and legacy engineering constraints are light. Inside a production codebase with historical debt, test baselines, and team conventions, the from-zero wish has far less room; the realistic use is prototyping and scaffolding new projects before humans take over maintenance. Impressive demo, narrow default: best-case behavior in the vendor's own environment, not a guaranteed floor everywhere.
6. Download and deployment: MIT license, weights fully open
The weights live on HuggingFace: deepseek-ai/DeepSeek-V4.1-Flash-1017. Since inline links are often not clickable in syndicated readers, here is the plain address again for copying: huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017.
ModelScope mirrors the full release; plain address: modelscope.cn/models/deepseek-ai/DeepSeek-V4.1-Flash-1017.
The license is MIT, continuing the September 10 policy: commercial use allowed, modification allowed, closed-source redistribution allowed, product integration allowed. Two reminders. First, at 510GB, plan the download route, ModelScope directly or a mirror in advance. Second, since the config files are identical to the September 10 build, an existing working deployment should run by swapping the weight directory, but validate the FP4 expert weight loading path on a test machine before production.
7. How this fits our other DeepSeek coverage
Our site already covers DeepSeek from several angles that do not overlap with this piece: the Ascend operator stack article covers kernel and hardware adaptation, the Cline coding agent article covers wiring DeepSeek into Cline, and the Harness desktop shell piece covers the desktop environment and onboarding. This article sticks to the 1017 weight update itself: parameter counts, benchmarks, cost, and the new capability.
FAQ
Q1: How does DeepSeek-V4.1-Flash-1017 relate to the September 10 V4.1-Flash?
A: It is a checkpoint update of the same model. Architecture, API name, and config files are unchanged; only the weights are new. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but retired, billed at Flash rates, and API callers need no code changes.
Q2: How large is the model, and how much space does a local deployment need?
A: Total weights are 763B (backbone 552B, Engram conditional memory 196B, roughly 15B of everything else), with activated parameters at 8B prefill and 16B decode. The checkpoint is 510 GB across 48 safetensors shards, experts in FP4 and the rest mixed precision. Plan disk around 510GB plus loading headroom.
Q3: Did API prices change?
A: No. Off-peak input is 0.15 dollars per million tokens on cache miss, 0.003 on cache hit, output 0.60 per million; peak hours double those. What changed is the bill: per-task cost fell because accuracy rose, with the DeepSWE figure dropping from about 0.50 to about 0.20 dollars (vendor chart reading).
Q4: Can the reported gains be trusted? Has anyone independent verified them?
A: All figures are vendor-reported and no third-party board has scored the 1017 build yet, so there is no independent verification. Opus 5 also still leads on Terminal-Bench 4.0 (51.8) and HLE (56.3) against the 1017 build's 42.3 and 42.6. Read it as a progress signal, not a claim of the top spot, and wait for independent runs before concluding.
Q5: Where can I download it, and what is the license? Any commercial restrictions?
A: HuggingFace and ModelScope host the full release; the plain address is in section 6: huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017. The license is MIT, the most permissive tier: commercial use, modification, and closed-source redistribution are all allowed with no extra terms.
Related reading
- DeepSeek Harness: 100K Stars in Two Days, "Everything Is a Plugin" — This Time DeepSeek Open-Sourced the Execution Layer, Not a Model: DeepSeek Harness: the vendor-built agent environment
- Cline on DeepSeek-V4-Pro: The 66K-Star Open-Source Coding Agent That Saves You Claude's Bill: Cline plus DeepSeek: a budget coding agent, assembled
- DeepSelect: DeepSeek's High-Performance TopK Kernel for DSA: DeepSeek DeepSelect open-sourced: the model routes itself