DeepSeek V4.1 Flash vs V4 Flash: Same Prompts, Same Hardware

Author: Prompt Engineer · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: DeepSeek V4.1 Flash vs V4 Flash: Same Prompts, Same Hardware · Channel: Prompt Engineer · 13:37 · Watch on YouTube

Topics: AI evaluation and benchmarks, Model inference and optimization, AI models, Local and open-source models

Summary

The video compares DeepSeek V4.1 Flash with the older V4 Flash using matched prompts, while noting that DeepSeek has retired the old API name and some routers silently serve the new model instead. In animation tests, V4.1 produced more detailed results and was more reliable on the challenging 3D donut, though it often used more tokens. The tests also cover image understanding, tool calls, a long-context retrieval task, and speed on shared hardware; V4.1 performed well, but results varied by host. The creator also critiques benchmark presentation, verbosity, pricing, and confusing model-version rollouts, concluding that V4.1 improves on the old Flash in the tested areas but does not beat V4 Pro on every knowledge benchmark.

Key points

DeepSeek retired V4 Flash on September 10 and redirected its old API name to V4.1 Flash; the older July 31 weights remain available through third-party hosts.
V4.1 Flash has a 1-million-token context window and a KV cache reported as one quarter the size of V4 Flash’s; the creator calculated a 3.95× reduction.
In the three animation tasks, V4.1 delivered a more detailed robot, won the donut test 2 times out of 3 versus the old model’s 1 out of 3, and tied on the rocket.
V4.1 was often more verbose: it used nearly twice as many tokens as the old model on the first robot test.
A vision test revealed that an OpenRouter request for the old model had silently fallen back to DeepSeek V4.1; with fallbacks disabled, the old model’s host did not support image input.
V4.1 correctly used parallel tool calls for stock-price lookups and currency conversions, and returned valid JSON on the first attempt.
Both models found a hidden sentence in a 182,279-token prompt; V4.1 started answering in about 4.5 seconds, compared with 8.5 seconds for the third-party-hosted old model.
Speed varied substantially across hosts, but on the same SiliconFlow host V4.1 ran at 187 tokens per second versus 87 for V4 Flash.
The creator notes that DeepSeek’s benchmark page omits its older models from the hero chart and that V4 Pro still leads V4.1 on some knowledge and reasoning results.
The video advises checking the provider and model fingerprint in API responses because model names, router fallbacks, and DeepSeek documentation may not match expectations.

Chapters

Start	Chapter	Summary
00:00	The retired model and test setup	The creator explains that the old V4 Flash API name now serves V4.1 and describes how they arranged a comparison using public weights and pinned hosts.
01:01	V4.1 design and benchmark claims	The video reviews V4.1 Flash’s architecture, cache size, benchmark results, and inconsistencies in DeepSeek’s announcements and documentation.
05:14	Three animation tests	Matched prompts test both models on a robot, a mathematically projected donut, and a looping rocket launch, comparing quality, reliability, and token use.
08:23	Vision, tools, effort, and context	The creator checks image input and router behavior, tests parallel tool calls and effort settings, and compares retrieval from a 182,279-token prompt.
10:28	Speed, pricing, and drawbacks	After discovering that different hardware distorted an initial speed comparison, the creator tests matched hosts and discusses costs, verbosity, and rollout issues.
13:05	Verdict	The creator summarizes where V4.1 Flash improved over V4 Flash, where V4 Pro still leads, and why users should verify the model they are actually calling.

Quotes

So, the model I asked for is gone. — 00:00

The old model got the donut right one time out of three. — 06:47

If you’re comparing models through a router, check the provider on every response. — 08:55

Yes. It got the hard animation right two times out of three, while the old model managed one. — 13:05

Concepts and entities

DeepSeek V4.1 Flash, DeepSeek V4 Flash, DeepSeek V4 Pro, OpenRouter, Artificial Analysis, KV cache, Mixture of Experts, context window

Full transcript (original language)

00:00 I asked Deepseek’s API for V4 Flash, the old model. The reply came back labeled Deepseek Flash with the exact same system fingerprint as the new one. So, the model I asked for is gone. Deepseek retired it on September 10th and quietly pointed the old name at V4.1. That makes a fair before and after test harder than it should be. Deepseek’s own table says V4.1 Flash beats their flagship V4 Pro. I wanted to check that against the old flash on the same prompts and Deepseek

00:30 no longer serves it. The weights are still public though. 28 hosts on Open Router still run V4 flash the July 31st build. So I pinned the old model to those hosts turn fallbacks off and ran both versions through the same tests. Three ASKI animations including a 3D donut that one model drew as a flat dotted line. vision tool calls a needle buried at 182,000 tokens and speed measured on the same hardware. Every number in this video came off my own screen except the ones I tell you belong

01:01 to DeepSeek. Now, what did they actually ship? V4.1 flash is a mixture of experts model with 552 billion parameters under an MIT license with the weights on hugging face the same day. The interesting part is the split deep sea calls it a causal encoder decoder. 40 layers cut in half. 20 layers read your prompt. 20 layers write the answer. Reading activates 8 billion parameters per token. Writing activates 16 billion. That asymmetry is aimed at agents. An

01:31 agent reads far more than it writes. File contents tool results the whole conversation again every turn. If reading costs half of what writing costs, the input heavy workloads get cheaper first. It also has vision built in from the start of pre-training. a 1 million token context window and 45 trillion training tokens. The part DeepS leads with on hogging face is the KV cache. That’s the memory the model keeps of everything it already read, so it doesn’t have to reread your whole prompt for every new word. V4 flash stored 3514

02:04 bytes of it per token. V4.1 stores 890. I divided those myself, 3.95 times smaller, which matches their one quarter claim. They got there with a few stack tricks. The cache is stored at four bits. Deeper layers reuse the indexes of earlier ones and the decoder builds its cache from the encoder’s output instead of keeping its own. The chart they published has one design choice worth noticing. The access starts at deepsec v1 389,000 bytes. So the step from v4 to

02:37 v4.1 is a sliver. You can barely see. The number is real. The picture just makes every recent model look like zero. At least for me, this is the most useful thing in the release because cash is what you pay for on long agent runs. Deepseek’s cash head price for flash is now 0.6 per million tokens at peak. Now, the benchmarks and the claim on the announcement page results ahead of flagship models including DeepSeek V4 Pro on agent encoding work. The table backs that up. Terminal Bench 2.19.6 6

03:09 against pros 87.9. Deepsw SWE 74.2 against 62.7. Terminal bench 3 against 11.8. Those gaps are big. The hero chart at the top of the page shows only four of the 19 rows. And it compares V4.1 against Kimmy, GLM, Opus 5, and GPT. Deepseek’s own V4 Pro and V4 flash columns are missing from the picture. They’re still in the full table underneath. And in that table, Pro still wins two rows. GPQA diamond 92.4 for pro

03:40 90.9 for V4.1. Humanity’s last exam is messier. V4.1’s headline number is 36.8 on the full set while pros 42.7 has an asterisk meaning text only subset. Compare like with like text only against text only and it’s 39.1 for v4.1 against 42.7. So pro keeps the knowledge and reasoning rose. In Fanis, they did leave a loss in the hero chart on terminal bench 3. Opus 5 scores 43.3 and V4.1

04:12 scores 30. They could have cut that bar and didn’t. Every one of these scores is deepseek measuring Deepseek using their own harness at maximum effort. So I checked artificial analysis which runs its own evals. They list V4.1 flash at max effort with an intelligence index of 46th of 113 models in its class at 209 tokens per second. They also flag it as very verbose 250 million output tokens to finish their eval suite against a

04:42 median of 130 million. Keep that number in mind because it shows up in my tests too. One more thing on the page doesn’t add up. The announcement says V4 Pro is being phased out and that from September 14th, every V4 Pro request routts to V4.1 Flash. The change log dated the same release says the opposite. In response to user demand, V4 Pro keeps running after September 14th. Billing unchanged. The pricing page agrees with the change log. So, the news page is out of date. And if you have V4 Pro in

05:14 production, the docs are the ones to trust. I called V4 Pro today and it answered with its own separate fingerprint. Now the fun part. I gave both models the exact same prompt. One self-contained HTML file, black background, a single block of green monospace text, plain JavaScript, 12 frames per second, looping forever, no libraries, no canvas, no images. Then three different animations, a walking robot, a spinning 3D donut with real projection math and shading, and a rocket launch with a countdown and a

05:44 scrolling starfield. V4.1 ran on DeepSeek’s API at max effort. V4 Flash ran on thirdparty hosts at its highest effort setting, one call each, and I recorded exactly what came back. Round one, the robot. The old model drew a simple boxy robot, square head, two dot eyes, antenna, legs, dead step. It works. It used 17,134 output tokens and took 49 seconds. V4.1 went further. Visor eyes that blink from open to shut. a chest panel, arms

06:16 hanging off the body, and feet that swap which one is planted. It also took 30,682 tokens and almost two minutes. Better robot nearly double the tokens that matches the verbosity artificial analysis flagged. Round two, the donut and this is the one that separated them. V4 Flash came back in 19 seconds with a page that runs without a single error. What it draws is a flat dotted line across the middle of the screen. The rotation math is in there. The projection collapses and you get a line

06:47 where a donut should be. One run can be bad luck. So I ran the old model twice more. The first retry thought for 259 seconds, burned through 131,000 tokens, hit the output cap, and returned an empty file. The second retry, 18 seconds, drew a small, properly shaded spinning donut. So the old model got the donut right one time out of three. V4.1 drew a full shaded donut on its first try. 12,683 tokens, 51 seconds. Light falls across

07:19 it with that ramp from dots to add signs and you can see the hole rotate. To be fair, I gave V4.1 the same two retries. The first one drew another good donut spinning on both axis. The second one crashed. A JavaScript arrow cannot set properties of null and a black screen. So, the final count on the donut is 2 out of three for V4.1, 1 out of three for the old model. The new model is more reliable on this and it still breaks sometimes. Round three, the rocket. And this one is a tie. Both models built a

07:50 starfield, a rocket on a pad, a countdown, flames, and a liftoff that loops. The old models version is tighter and faster. V4.1 has a wider launchpad and a cleaner exhaust trail. Neither broke. V4.1 used 34,000 tokens and the old model used 43,000. So on this round, the new model was the more economical one. Two clear wins for the new model on detail and on the hard math. One tie and a pattern. V4.1 thinks longer and fails less. Vision is the one feature. The old

08:23 model can do it all. And testing that turned up something I didn’t expect. I sent V4.1 Deepseek’s own KV cache chart and asked for every model name and number. Then the ratio. It read all four values exactly 389,120 down to 890 and computed 3.958 seconds. Then I sent the same image to the old V4 flash on open router and it answered perfectly too. That shouldn’t be possible for a text only model. So I looked at who served the request. Open

08:55 router had quietly routed it to DeepSseek’s own endpoint and the model field in the response said Deepseek v4.1 flash. I asked for the old model and got the new one again with fallbacks turned off. The honest answer came back. No endpoints found that support image input. If you’re comparing models through a router, check the provider on every response. Three quick feature checks on V4.1 tool calling. I gave it two fake tools, a stock price lookup and a currency converter, and asked what 10 shares each of Nvidia and Apple would

09:26 cost in rupees. It called both price lookups in parallel, fed the results into two parallel conversions, and then answered 1 lak 78,840 rupees for Nvidia, 28,900 for Apple. I checked the arithmetic. It’s correct. And it added a note that live quotes move. JSON mode returned valid JSON first time. The effort levels didn’t behave the way I assumed on a counting problem with a brute forced answer of 432. All six runs got it

09:58 right. Low effort used between 1,00 and 2,100 reasoning tokens. Max used 2,100 to 2,800. High actually used the fewest on one run on a problem this size. low effort gave the same answer for less. So, I’d start there and only raise it when answers go wrong. Then, context. I build a hay stack of 2,000 paragraphs of filler, 182,279 tokens, and hid one sentence 63% of the way in. The secret launch code for the

10:28 orange submarine is Tangerine 4471. Both models found it. V4.1 started answering in 4 1/2 seconds, and the old model on a third party host took 8 and 1/2. The prompt cost V4.1 about 3 cents at off peak rates. One needle at one depth as a sanity check and both passed it. Speed is where I nearly got fooled. My first run had the old model at 295 tokens per second and V4.1 at 190. That looks like the new model is slower. The old model

10:59 was running on a host called RA and V4.1 on deep servers. I was comparing hardware. So I pinned both models to the same host. Three runs each fall backs off on Silicon Flow. The old model ran at 87 tokens per second and V4.1 at 187, more than double. On Novita, 139 against 189 on Parasel. The results were all over the place. V4.1 swung between 15 and 141 across three runs. So, I’m

11:31 throwing that host out. On DeepSync’s own API, V4.1 held 190 tokens per second across three runs and V4 Pro managed 73 for comparison. Artificial analysis measured 209 price. V4.1 Flash costs $120 per million output tokens at peak and 60 off peak. V4 Pro costs $3.96 at peak. Input on a cash miss as 30 cents at peak and a cash hit as 6 of a cent. Off peak as half price and off peak

12:02 includes all of the weekend. This entire video’s testing. Three animation rounds with retries, vision tools, effort sweeps, a 182,000 token needle and nine speedruns on Deep Seeks API cost me about 9. What’s wrong with it? first verbosity. It used almost double the tokens of the old model on the robot and artificial analysis measured it at nearly twice the median. The per token price is low and a model that thinks for 30,000 tokens on a cartoon robot spends

12:33 it quickly. Second, the benchmark presentation. The hero chart leaves out Deepseek’s own older models and the HLE headline compares a full set score against texton scores. Pro still wins knowledge and reasoning. Third, the roll out. Old model names silently redirect. The news page and change log disagree about V4 Pro and routers fall back to the new model without telling you. If you pin a model version in production, you now have to verify. It’s really the version you pinned. So, is V4.1 Flash better than the model it replaced on

13:05 everything I could measure? Yes. It got the hard animation right two times out of three, while the old model managed one. on the same hardware. It ran more than twice as fast on one host and a third faster on another. It reads images and it calls tools cleanly in parallel. Is it better than V4 Pro on agents coding and speed? Yes, at under a third of the output price. On knowledge questions, Deepseek’s own table says Pro still leads and it’s open weights under MIT. The model you call today can be retired out from under you because you

13:37 can download it. If you’re interested in reading further, I link to this main page in the description below. One question for the comments. Have you caught a router or an API serving you a different model than the one you asked for? If this was useful, a like and a subscribe helps the channel a lot. Thanks for watching.