What Is DeepSeek V4.1 Flash?
In plain English, DeepSeek V4.1 Flash appears to be a test build for the next iteration of DeepSeek's speed-oriented V4 model. It is newer than DeepSeek V4 Flash 0731, but “newer” does not yet mean “officially released.”
The internal beta notice was reproduced by developers on Hacker News and Reddit. Both copies describe it as an intermediate version and list the same temporary model ID, beta pricing rule, and concurrency cap. The public DeepSeek documentation, by contrast, still lists only deepseek-v4-flash, deepseek-v4-pro, and deepseek-v4-flash-vision-exp.
That gives us a useful evidence boundary:
| Item |
What is known on September 9, 2026 |
| Developer |
DeepSeek |
| Status |
Intermediate internal beta; not a documented GA release |
| Temporary model ID |
deepseek-v4.1-flash-expires-on-0910 |
| Reported test start |
September 8, 2026 |
| Temporary ID expiry |
September 10, 2026 |
| Reported input |
Native multimodal support; early users have tested text and images |
| Output |
Text |
| Beta concurrency |
20 requests per account |
| Beta billing |
Same rate card as DeepSeek V4 Flash |
| Context and maximum output |
Not published for V4.1 Flash |
| Model weights and license |
Not published for V4.1 Flash |
| GPT Proto availability |
Not available yet |
The short version: DeepSeek V4.1 Flash is real enough to test, but not documented enough to specify with production confidence.
DeepSeek V4.1 Flash Release Status and Timeline
V4.1 Flash makes more sense when placed beside the releases that came before it.
| Date |
DeepSeek V4 milestone |
What changed |
| April 24, 2026 |
DeepSeek V4 Preview |
V4 Pro and V4 Flash arrived with a 1M-token context window and agent-focused improvements. |
| July 31, 2026 |
V4 Flash 0731 |
The updated Flash API entered public beta with stronger coding-agent results and Responses API support. |
| August 13, 2026 |
V4 Pro 0813 |
V4 Pro reached general availability with updated agent capabilities and additional reasoning-effort controls. |
| August 21, 2026 |
V4 Flash Vision Exp |
DeepSeek added an experimental text-and-image version of Flash. |
| September 8, 2026 |
V4.1 Flash internal beta |
A temporary intermediate build appeared with a new model ID and reported native multimodality. |
| September 10, 2026 |
Temporary endpoint expiry |
The beta ID indicates that this particular test endpoint will expire. It is not a confirmed GA date. |
DeepSeek's V4 Preview announcement describes V4 Flash as the faster, less expensive member of the family and V4 Pro as the higher-capability tier for harder agent tasks. The Vision Exp release then added image understanding to the Flash line.
V4.1 Flash appears to be the next step in that sequence. Still, it would be a mistake to turn the September 10 expiry into a release prediction. DeepSeek could replace the test build with another checkpoint, announce a stable model later, or withdraw it for more work. No public source gives a final date.
What Is New in DeepSeek V4.1 Flash?
The beta notice makes four claims: a new architecture, native multimodality, greater capability, and better speed and cost efficiency. Only part of that can be examined today.
| Claim |
Current evidence |
What remains unknown |
| New model architecture |
Stated in the shared beta notice |
Architecture diagram, parameter count, active parameters, attention design, and training recipe |
| Native multimodal support |
Stated in the notice; developers have submitted image-based tasks |
Exact supported media types, image limits, tokenization rules, and whether video or audio will be supported |
| Stronger capabilities |
Stated in the notice |
Official evaluations and independent head-to-head results |
| Faster inference |
Multiple early API users report high decoding rates |
Stable throughput under normal traffic, latency distribution, and performance after GA |
| Lower cost |
Stated as a design goal |
Final token prices and cost per successfully completed task |
The phrase “native multimodal” deserves attention. DeepSeek V4 Flash is text-only, while V4 Flash Vision Exp is an experimental variant that accepts text and images. DeepSeek says Vision Exp matches V4 Flash on text tasks and converts each image into up to 384 billable tokens. V4.1 Flash may combine the two paths more directly, but DeepSeek has not explained how its multimodal components are trained or connected.
So the upgrade is potentially meaningful, especially for agents that must inspect screenshots before acting. The mechanism is still a blank box.
How Fast Is DeepSeek V4.1 Flash?
Speed is the clearest early signal—and the easiest one to exaggerate.
Developer Ivan Fioravanti reported an average decoding rate of about 350 tokens per second during the beta. His post, reproduced in the NVIDIA developer forum, also described the model as clearly experimental and under heavy load. Other users have reported results in roughly the 300–500 tokens-per-second range, but none of these figures is an official service guarantee.
One community coding test compared V4.1 Flash with V4 Flash Vision Exp using the same app-building prompt and agent setup:
| Result from the single community run |
V4.1 Flash |
V4 Flash Vision Exp |
| Total duration |
18m 49s |
30m 11s |
| Total token usage |
11.59M |
20.31M |
| Output tokens |
126K |
154K |
| Reported model time |
7m 48s |
25m 10s |
| Generation speed |
361 tokens/s |
120 tokens/s |
| Cache hit rate |
99.5% |
99.7% |
On the surface, V4.1 Flash finished about 38% sooner and used about 43% fewer total tokens. But the tester also noted that the V4.1 session still had one task in progress and another pending, whereas Vision Exp had completed all 11 tasks. This was a useful early coding report, not a controlled benchmark.
For another reference point, Artificial Analysis currently measures the documented V4 Flash 0731 API at 125.7 output tokens per second with a 0.92-second time to first token. That does not prove V4.1 is nearly three times faster. The beta may be running on a different hardware allocation, under different traffic, with different prompts and output lengths.
This distinction matters because output speed is only one part of an agent's latency. A useful evaluation also measures time to first token, reasoning time, tool execution, retry count, total tokens, and whether the task was actually completed. A model that emits tokens twice as quickly but requires another attempt may cost more time overall.
DeepSeek V4.1 Flash Pricing
The shared beta notice says DeepSeek V4.1 Flash uses the same pricing as deepseek-v4-flash during testing. It does not establish the final V4.1 Flash price.
DeepSeek currently uses peak and off-peak pricing for V4 Flash. Its official model and pricing table lists the following rates per one million tokens:
| Token type |
Off-peak price |
Peak price |
| Cached input |
¥0.05 |
¥0.10 |
| Uncached input |
¥1.50 |
¥3.00 |
| Output |
¥4.50 |
¥9.00 |
Peak periods are 09:00–12:00 and 14:00–18:00 Beijing time, Monday through Friday. All other periods use the lower rate. The temporary V4.1 beta is also capped at 20 concurrent requests per account, compared with 2,500 listed for the normal V4 Flash API.
“Lower cost” therefore needs two definitions. The beta's token rate is reportedly unchanged, so DeepSeek is not giving users a new per-token discount during this test. It may instead mean lower internal inference cost, fewer tokens needed to finish a task, or a lower future price. Only the first part—the temporary rate card—is known.
GPT Proto has not set a V4.1 Flash price because the model is not yet available on the platform. Developers who need a stable DeepSeek endpoint today can review the existing DeepSeek V4 Flash API instead.
DeepSeek V4.1 Flash vs V4 Flash, Vision Exp, and V4 Pro
The comparison below deliberately leaves V4.1 fields blank where DeepSeek has not published a specification.
| Model |
Release status |
Input |
Context / max output |
Public model card |
Best current role |
| DeepSeek V4.1 Flash |
Temporary internal beta |
Reported native multimodal input |
Not published |
No |
Short evaluation runs |
| DeepSeek V4 Flash 0731 |
Public API beta |
Text |
1M / 384K |
Yes, MIT-licensed weights |
Fast text coding and agents |
| DeepSeek V4 Flash Vision Exp |
Experimental public API |
Text and images |
1M / 384K |
Yes |
Visual agents and screenshot-based tasks |
| DeepSeek V4 Pro 0813 |
General availability |
Text |
1M / 384K |
Yes, MIT-licensed weights |
Harder reasoning and agent workflows |
DeepSeek V4.1 Flash vs DeepSeek V4 Flash
The expected upgrade is straightforward: retain the speed-oriented Flash position while adding native multimodality and improving decoding speed. The existing DeepSeek V4 Flash has a documented 1M context window, a maximum 384K output, published weights, and a stable model ID. V4.1 has none of those guarantees yet.
Choose V4 Flash for an integration that must keep working after September 10. Test V4.1 only if learning about the next version is more important than endpoint continuity.
DeepSeek V4.1 Flash vs V4 Flash Vision Exp
Vision Exp is the better-known visual model today. It accepts text and image input through Chat Completions, Messages, and Responses formats, and DeepSeek documents base64, URL, and Files API image inputs. V4.1 Flash may become the cleaner long-term multimodal successor, but that is an interpretation of the beta direction, not a confirmed migration plan.
DeepSeek V4.1 Flash vs DeepSeek V4 Pro
V4 Pro remains DeepSeek's documented higher tier. DeepSeek's own V4 materials say Flash can approach Pro on simpler agent tasks but trails it on harder work. No V4.1 benchmark currently proves that the new Flash build closes that gap.
Use DeepSeek V4 Pro when difficult task completion matters more than the lowest latency. A fast beta is not automatically a Pro replacement.
DeepSeek V4.1 Flash vs GLM-5.3 Flash and Claude Fable 5.1
V4.1 Flash is also being discussed beside newer multimodal and agent-focused models. The fair comparison is not a single leaderboard score. It is the combination of maturity, modality, latency, price, and failure cost.
| Model |
Product position |
Published context |
Multimodal input |
Published weights |
Availability on GPT Proto |
| DeepSeek V4.1 Flash |
Speed-oriented intermediate beta |
Unknown |
Reported native multimodal |
No |
Not yet |
| GLM-5.3 Flash |
Low-cost multimodal coding and agent model |
1M |
Text, image, video, and files |
MIT |
Available |
| Claude Fable 5.1 |
Premium long-horizon reasoning model |
1M |
Text and images |
No |
Available |
| GPT-6 Astra |
Premium end-to-end reasoning and computer-use model |
1.05M |
Text, images, and documents |
No |
Available |
DeepSeek V4.1 Flash vs GLM-5.3 Flash
GLM-5.3 Flash is the closest established comparison because it is also positioned around inexpensive multimodal coding and agents. Its official model card documents 320 billion total parameters, 18 billion active parameters, native multimodality, and an MIT license. V4.1 Flash has not disclosed equivalent specifications.
Artificial Analysis currently records 73.1 output tokens per second for GLM-5.3 Flash and 125.7 for the older DeepSeek V4 Flash 0731 in its matched comparison. That may make the early 300-plus V4.1 reports plausible, but it still does not produce a valid V4.1-versus-GLM result. GLM is the safer choice for a deployment today; V4.1 is the more interesting short test.
DeepSeek V4.1 Flash vs Claude Fable 5.1
These models occupy different price and risk tiers. Anthropic positions Claude Fable 5.1 for demanding reasoning and long-horizon agentic work. It has a 1M context window, a maximum 128K output, and a documented API price of $10 per million input tokens and $50 per million output tokens. GPT Proto currently lists it at $9 and $45.
V4.1 Flash is likely to be much less expensive and faster at token generation. Fable 5.1 offers documented behavior, a stable model ID, and a clearer fit for tasks that run for hours. If one failed agent run costs more than the model bill, maturity can matter more than token price.
GPT-6 Astra belongs in the same high-end group. OpenAI documents a 1.05M context window, a maximum 128K output, and support for complex reasoning, coding, computer use, research, and document creation. It is a mature alternative when an agent must operate tools and interfaces now rather than wait for V4.1's final form.
DeepSeek V4.1 Flash for Coding, Agents, and Work
The most credible V4.1 Flash applications follow from its reported combination of high decoding speed and image input:
Frontend coding that uses screenshots to identify layout defects and check revisions.
Coding agents that inspect repository files, run tools, and return large patches.
Document workflows that mix text instructions with charts, scanned pages, or interface captures.
GUI agents that must interpret what is visible before choosing the next action.
High-volume internal evaluations where throughput matters and a temporary endpoint is acceptable.
These are evaluation targets, not verified promises. DeepSeek has not published V4.1 results for coding-agent benchmarks, visual reasoning, tool calling, or long-context retrieval. Its 20-request concurrency cap also makes the beta unsuitable as the only endpoint behind a production service.
What Should Developers Test Before Production?
A speed test alone is too shallow. I would use a fixed task set and record both model behavior and system behavior:
Run the same repository task at least several times and count accepted completions, not just tokens per second.
Verify that every tool call either receives a result or ends with a clear error.
Add an idle timeout for streaming and confirm that the client receives a final completion signal.
Test multi-turn Responses API calls and preserve the relationship between tool outputs and their call_id values.
Watch for repeated reasoning that consumes tokens without advancing the task.
Test mixed text-and-image requests with the exact image sizes and formats used in production.
Record cache hits, uncached input, output tokens, retries, and cost per accepted result.
Keep a stable fallback model available before sending real customer traffic.
Why these checks? Users of earlier V4 endpoints have reported premature stops during tool loops, SSE streams that remain open after output stops, tool-output pairing errors in the Responses API, and repeating reasoning loops. These reports concern V4 Flash or Vision Exp, not V4.1 Flash. They are a regression checklist—not evidence that the beta has the same problems.
Can You Use DeepSeek V4.1 Flash Now?
At the time of this update, some existing DeepSeek API users report that they can keep the normal DeepSeek base URL and switch the model name to deepseek-v4.1-flash-expires-on-0910. The endpoint is temporary, has a 20-request concurrency limit, and may disappear after September 10.
This page does not include a copy-and-paste API example for that reason. A code block built around an expiring beta ID would become misleading almost immediately.
DeepSeek V4.1 Flash is not yet available through GPT Proto. GPT Proto is evaluating the model and will update this guide after a stable model ID, final pricing, supported request formats, and production availability have been verified.
If you need an endpoint now, the current options are:
Is DeepSeek V4.1 Flash Worth Trying?
Yes—for evaluation. Not yet for production.
V4.1 Flash is worth testing if you already use DeepSeek's API, want to evaluate visual coding or multimodal agents, and can finish the experiment before the temporary endpoint expires. Early decoding-speed reports are unusual enough to merit attention.
Wait if you need a durable model ID, a published context limit, open weights, an SLA, repeatable independent benchmarks, or confidence that tool-based sessions will complete. Those are not minor documentation gaps. They are the difference between an interesting checkpoint and a production model.
Final Takeaway
DeepSeek V4.1 Flash is best understood as a preview of DeepSeek's next Flash direction: native multimodal input, very high reported decoding speed, and an emphasis on lower inference cost. The early evidence is promising, but thin. Its temporary ID expires on September 10, the beta allows only 20 concurrent requests per account, and core specifications remain unpublished.
Test it if the learning is useful. Keep it out of the critical path until DeepSeek publishes the final model and independent results can separate architectural progress from temporary infrastructure conditions.