Appearance
What Is DeepSeek V4.1 Flash? The Model That Topped OpenRouter's Weekly Usage
On October 26, OpenRouter published its token table for the week of October 19 to 25. The model at the top was not GPT-6 Luna, not Claude Opus 5.5, not Gemini 4 Argon. It was a checkpoint DeepSeek had pushed out nine days earlier: DeepSeek V4.1-Flash-1017, at 33.6 trillion tokens for the week, up 801% week over week.
That result is easy to misread both ways. Read it as proof that a budget model has beaten the frontier labs and you draw the wrong conclusion. Read it as a leaderboard quirk and you miss the real signal: developers pick on output-token price, they migrate in days rather than quarters, and they will run serious agent workloads on a model that costs cents per million output tokens.
This breakdown covers what V4.1 Flash is, where it sits in DeepSeek's lineup, how to read the 33.6 trillion figure without inflating it, how it prices against its closest rivals, and where it is the wrong tool. Teams that want to test it next to GPT, Claude and Gemini without stacking subscriptions can do that inside Nolvia from one browser tab.
Table of Contents
- What DeepSeek V4.1 Flash Actually Is
- Where Flash Sits in the DeepSeek Lineup
- Reading the 33.6 Trillion Token Number
- Why Usage Jumped 801% in One Week
- DeepSeek V4.1 Flash vs DeepSeek V4 vs Kimi K3 vs GLM-5.3 Flash
- What the 1017 Checkpoint Changed
- What This Means for Developers
- Accessing Chinese Models Reliably From Outside China
- FAQs
- Related Articles
What DeepSeek V4.1 Flash Actually Is
DeepSeek released V4.1 Flash on September 10, 2026 as the smallest model in a new architecture family, built for faster inference, more throughput, and a structure that scales up to larger parameter counts later.
The architecture is the interesting part. It is a sparse mixture-of-experts model with a 552B-parameter backbone that does not spend compute symmetrically. Roughly 8B parameters are active per token during prefill, when the model reads the input, and 16B during decode, when it writes the answer. DeepSeek calls the design a causal encoder-decoder, and the split is deliberate, since decoding is where reasoning-heavy generation happens. Compressed KV caching cuts cache memory to about a quarter of the previous Flash generation, which is what makes long agent sessions affordable at all.
Context length is 1,048,576 tokens. Vision is native, trained jointly with text from pre-training rather than bolted on afterward as in the experimental V4 Flash Vision Exp release from August. Weights ship under an MIT license, which permits commercial use.
The version that topped OpenRouter is a refresh, not a new model. On October 17, DeepSeek upgraded V4.1 Flash to DeepSeek-V4.1-Flash-1017. The model name deepseek-flash, the price and the endpoints all stayed put. What changed was post-training: better reasoning, coding, automation and vision performance, plus search and fetch endpoints for pulling live web content mid-task.
Where Flash Sits in the DeepSeek Lineup
DeepSeek runs a two-tier lineup, and the names are literal. Pro is the capability tier. Flash is the cost tier, built to be cheap enough that you stop thinking about per-call cost.
That lineup is currently upside down. V4 Pro reached general availability on August 13 at $0.66 per million input tokens and $1.98 per million output tokens off-peak, doubling to $1.32 and $3.96 at peak. V4.1 Flash sits at $0.15 and $0.60 off-peak, or $0.30 and $1.20 at peak. DeepSeek's own positioning is that V4.1 Flash exceeds V4 Pro on performance, speed and task completion time while charging roughly a quarter of the output price. When the cheaper tier outruns the flagship on the vendor's numbers, the flagship stops being the default.
The peak and off-peak split changes real bills. Peak hours are Beijing time, Monday to Friday, 09:00 to 12:00 and 14:00 to 18:00; everything else, including weekends, bills at half rate. A batch job scheduled overnight from Europe or North America lands almost entirely off-peak, which makes scheduling the easiest cost lever available.
Cached input is the second lever: cache reads bill at $0.003 per million tokens off-peak. In a coding agent that re-sends the same repository every turn, cached input is most of the input, so effective input cost collapses far below the headline $0.15.
Reading the 33.6 Trillion Token Number
OpenRouter routes requests to many providers behind one account, and its weekly token table has become the closest public meter of developer demand. For October 19 to 25 it recorded 188 trillion tokens across all models, up 14.63% week over week. V4.1-Flash-1017 accounted for 33.6 trillion, roughly 18% of the platform's traffic in one week.
The rest of the table shows the shape of the week. The September 10 version was second at 18 trillion tokens, down 45%. MiniMax M3.1 Flash was third at 13.7 trillion, Z.ai's GLM-5.3 Flash fourth at 13.2 trillion, and Tencent's Hy4 preview fifth at 8.97 trillion. All five top slots went to Chinese models, which took 93.54 trillion against 22.47 trillion for US models, the twenty-sixth straight week Chinese models led here.
The caveats matter more than the headline. OpenRouter counts prompt and completion tokens including reasoning tokens, and only traffic through its own routing layer. Vendor-owned APIs, cloud platforms, private deployments and consumer apps sit outside the measurement. The country comparison therefore describes aggregator traffic, not total industry usage, and the figure is not revenue or user count.
What it does say is narrower. On a platform where users pay per token from their own balance, 33.6 trillion tokens is a lot of deliberate choices, and the pattern of one version up 801% while its predecessor drops 45% says switching costs between checkpoints here are close to zero.
Why Usage Jumped 801% in One Week
The 801% is real, but it is not organic growth in the usual sense. Three things stacked.
Version migration came first. DeepSeek moved its endpoints to the 1017 checkpoint on October 17 and users changed nothing, because deepseek-flash kept pointing at the same endpoint at the same price. Traffic previously counted toward the September 10 model started counting toward the new one. The old version's 45% decline and the new version's gain are the same users in motion, so much of the 801% is cannibalization, not new demand.
The refresh was also a genuine capability jump at a flat price. On DeepSeek's published numbers, Terminal-Bench 4.0 rose from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, DeepSWE v1.1 from 74.2 to 77.0, and Humanity's Last Exam from 36.8 to 42.6. DeepSeek called it a Pareto improvement on performance versus cost: better scores, shorter runtimes, shorter outputs. Raising capability without raising price is the cheapest growth strategy in this market.
The third factor is structural. Open weights plus many hosts equals price pressure: any provider can serve an MIT-licensed model, and they compete down toward hardware cost. DeepSeek open-sourced its DeepInfer serving framework on October 20. Z.ai's GLM-5.3 Flash activates 18B of 320B parameters, and MiniMax launched M3.1 Flash claiming triple the inference speed at 70% lower cost. Competition has moved from parameter count to cost per unit of useful work.
DeepSeek V4.1 Flash vs DeepSeek V4 vs Kimi K3 vs GLM-5.3 Flash
The four models worth comparing are not the biggest names, but the ones a cost-conscious team would shortlist this month. Prices are per million tokens at list rates in late October 2026, and provider prices vary, so read them as a floor.
| Model | Input price | Output price | Context | Active / total parameters | Open weights |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash (1017) | $0.15 off-peak / $0.30 peak | $0.60 off-peak / $1.20 peak | 1.05M | 8B prefill, 16B decode / 552B | Yes (MIT) |
| DeepSeek V4 Pro (0813 GA) | $0.66 off-peak / $1.32 peak | $1.98 off-peak / $3.96 peak | Not listed | Not disclosed | — |
| Kimi K3 | $3.00 | $15.00 | 1.05M | 2.8T total (MoE) | Yes |
| GLM-5.3 Flash | From $0.03, varies by provider | From $0.10 | 1M | 18B / 320B | Yes |
Two things stand out. Input prices sit within an order of magnitude of each other while output prices do not. V4.1 Flash at $0.60 output against Kimi K3 at $15 is a 25x gap, and in agent loops where output tokens dominate the bill, that gap is the entire decision. Kimi K3 is not competing on price; it is a 2.8T-parameter open-weight reasoning model for long-horizon repository work.
The naming needs care too. Z.ai's GLM-5.3 Flash is the tier that appeared fourth on the OpenRouter table. Full GLM-5.3 is a different product at a different price, listed at ¥8 input and ¥28 output per million tokens on Alibaba's Model Studio. Comparing V4.1 Flash against a Flash-tier rival is honest; comparing it against a flagship and calling it cheaper is not.
What the 1017 Checkpoint Changed
DeepSeek published the September 10 baseline and the October 17 refresh against the same benchmark names, so the delta is visible.
| Benchmark | V4.1 Flash (Sep 10) | V4.1-Flash-1017 (Oct 17) |
|---|---|---|
| Humanity's Last Exam (no tools) | 36.8 | 42.6 |
| Terminal-Bench 4.0 | 31.2 | 42.3 |
| DeepSWE v1.1 | 74.2 | 77.0 |
| ProgramBench | 20.3 | 29.1 |
| NL2Repo-Bench | 65.4 | 68.2 |
| Agents' Last Exam | 31.8 | 35.4 |
These are vendor-reported and harnesses differ by benchmark, so read them as direction rather than absolute standing; Terminal-Bench 4.0 ran on the Claude Code harness, DeepSWE on Mini-SWE-V2. What the table does establish is where the refresh went: agentic coding and tool use, the tasks that burn output tokens.
What This Means for Developers
Reach for V4.1 Flash when output tokens dominate your cost structure and the task is verifiable. Batch code generation, repository-scale refactors, log triage, extraction pipelines, agent loops that retry until a check passes, long-context document analysis. In each case you can measure whether the output is right, you pay mostly for output, and a 25x price gap against a premium reasoning model changes what is worth automating: work too expensive to run becomes affordable when per-task cost drops from dollars to cents.
Skip it when the task cannot be checked and being wrong is expensive: hard one-shot reasoning, novel research synthesis, legal or financial analysis where a confident wrong answer is worse than none, anything needing a contractual relationship or data residency. DeepSeek's own numbers still put the frontier closed models ahead on the hardest reasoning evaluations.
The operational risk is churn. A model that gained 801% in a week because a checkpoint shipped can be displaced just as fast; the previous version losing 45% the same week is proof. Pin explicit checkpoint identifiers in production, keep your own eval set, and re-run it whenever a provider announces a refresh.
That is also the practical case for keeping several models in reach: Nolvia puts DeepSeek, GPT, Claude, Gemini and 40+ models behind one login, turning side-by-side evaluation into a ten-minute job.
Accessing Chinese Models Reliably From Outside China
Choosing between these models is the easy part. Reaching them is not. DeepSeek, Moonshot and Z.ai each run their own console, billing, rate limits and availability profile, and signing up from outside the region adds payment friction. Three vendors means three accounts and three sets of API behaviour to learn.
Nolvia is built for that gap. It is a browser-based multi-model workspace: one subscription, one interface, model switching inside a conversation, no per-vendor onboarding. Plans run at $15/mo for 45K points, $30/mo for 100K points and $60/mo for 200K points, and point consumption scales with the model you pick, so a Flash-tier model stretches much further than a frontier one.
One boundary worth stating: Nolvia is a web workspace, not an inference reseller. If your production pipeline needs programmatic endpoints and per-token billing, go direct to the provider. If you need a place to actually use these models, a workspace gets you there faster.
Try Every Major AI Model in One Workspace
Nolvia gives you GPT, Claude, Gemini, Grok and 40+ models in a single subscription — no API keys, no juggling five tabs.
Start Free on Nolvia