We will adjust the pricing for the Flash series effective from 12:00 Beijing Time on September 10, 2026. During off-peak hours, the unit price will be $0.003 for input cache hits, $0.15 for input cache misses, and $0.6 for output. Peak-hour prices will be double the off-peak rates. Please plan your usage accordingly.
But, the web ui chat version of flash has very poor language following abilities in my experience:
You may ask it something in English, and get a thinking chain in Chinese with an answer in Chinese, or an English thinking chain and an English answer. Using the retry button on the same question has a 50/50 chance of any of those results.
Sometimes, asking something in English, but where information are mostly in another language may make the answer in the language where data has been found. The other day, I asked something about a local German thing, in English, and I got an answer in German instead. It’s as if all the language data stirred it away from the language of the user’s question.
Hallucinations you can't fix. Gemini is a bit worse there than DeepSeek, but there's not much research on how to fix that. The only one is the CaMeL paper by Google, where you tag every prompt and result and then for every assistant response or tool call you first check where it got that data and error if you notice fabrication. This one is really annoying to implement.
With larger models the fabrication starts when the context grows or if you have too many tools, for flash models it's much earlier. We use the flash models for repetitive agentic tasks, where the prompt defines clearly what to do and how. The whole run is about 4-5 steps typically, and context size stays in the comfort zone.
I initially thought it was a trick, that using Chinese chars is somehow more info dense and it saves tokens to 'think' in Chinese.
But later on it became more erratic. I still wonder if token reduction would work that way.
Interesting though, when I ask questions in German or my native language, I rarely get Chinese answers. Looks like English is most affected.
API never answers in Chinese.
Hasn't happened in a while, last time was when I was testing fable 5 in june.
Please don't do this kind of thing. If a user has validated a workflow on V4 Pro, they might not want to suddenly start testing it in production on V4.1 Flash. Instead, keep V4 Pro around but deprecated for a defined period of time, then remove it.
At least as open weights models, it's possible to use something like Together.ai or OpenRouter to run the V4 Pro model as long as other providers keep it up.
But I agree with you. I have a dumb workflow that worked well with v4-flash-0731 and I suspect is directing to a newer model that now breaks it.
Where Gemini still wins is non-text input what Deepseek cannot do, yet, and Deepseek Flash has this thing of cheaper models where a failing tool call can derail your agent to a retry loop if you're not careful on instructions in the error message.
If they fix and make the tool calls to work better in non-optimal situations, it's much easier to switch from Gemini without a few weeks of evals and bugfixing.
I make vaporware that doesnt do shit reliably and this chinese crap spouts plausible demos and spam calls more cheaply than the competition saaar
I don't need a model that can invent new mathematics. I need something that is fast, cheap, and consistent. Give me that and I can build and scale.
Yeah, that’s basically an industry-wide scam.
In any case old rules apply: if privacy is a concern don't share the data. I share all my work-related code because it's worthless, but I don't and would never share company business and process details, access to production/user data, etc.
Meanwhile I know of people connecting all the kind of MCPs for datadog/sentry/jira/concluce/production databases to their harnessess..lol.
Works fantastic. Glad there is a more 'uncensored' thing to fall back to when the frontier folk are too sensitive.
Incredible good value and product they have built.
https://www.geeky-gadgets.com/deepseek-v4-1-flash-review/
I hope some of those speed increases will make it to production.
I wonder if this comes from using the bad architecture scaled up (and it hits some limits) or if this is a data problem (undertrained? bad data? bad pre-processing using smaller models?)...
V4 flash and V4 pro feel very similar, which would make sense if they were pre-trained on largely the same corpus.
All that would suggest to me is that V4 Flash is capable of absorbing the data they’re throwing at it, and we’re still nowhere near the data limits of their larger 1.6T model
The announcement specifically says 4.1 Pro will be released in the future.
For all intents and purposes, "low" is pretty much the same as turning reasoning off, and "high" is similar to "max". "High/max" performs way too much reasoning, takes forever, and causes costs to balloon. They need a proper "medium" setting.
I get it that they're probably focused on pushing performance right now, but the ergonomics of the model aren't great.
If I'd carefully tested and optimized prompts against Pro I wouldn't be keen on this particular news. I feel like API model providers should lean towards not swapping out models on their paying customers, no matter how much "better" the new model is meant to be.
I'm making my own quants, though the Vision-Exp version is outdated and won't work on llama.cpp master branch (I built it before llama added support):
- https://huggingface.co/tarruda/DeepSeek-V4-Flash-0731-GGUF
- https://huggingface.co/tarruda/DeepSeek-V4-Flash-Vision-Exp-...
For the Vision-exp version, I also ran perplexity + KLD against the original MXFP4. Seems quite OK: https://huggingface.co/tarruda/DeepSeek-V4-Flash-Vision-Exp-...
I already have new GGUFs but haven't uploaded yet. If you want Vision-Exp, maybe use bartowski or unsloth's GGUFs.
Side note:
As an alternative to deepseek v4, you might want to give it a shot at qwen 3.8 flash next. I have IQ4_NL GGUFs that can be loaded fully into 128G, or Q5_K GGUFs that can offload the PLE to disk (use --load-mode none --lazy-mode on for that): https://huggingface.co/tarruda/Qwen3.8-Flash-Next-GGUF.
llama.cpp master is still somewhat bad in Qwen 3.8 next performance, but I was able to achieve 40tps tg and 600 tps pp on my private branch.
While V4.1 Flash performance and cost looks promising this auto re-routing sounds concerning
Its cost is now 1/10th per token, and 1/5th per task.
Basically they have shitty hardware so they have to do a lot of optimization. Think of it like replacing an O(n) algorithm with O(log n).
Anthropic / Open AI think the best path is the most intelligent models deepseek is more focused on tok/$
hope deepseek makes me change my setup again
Fairly excited for the v4.1 launch. Input cache hit prices have been halved, which looks nice.
Looks like the new model can be used if summoned via the API but the API won't list it.
We will adjust the pricing for the Flash series effective from 12:00 Beijing Time on September 10, 2026. During off-peak hours, the unit price will be $0.003 for input cache hits, $0.15 for input cache misses, and $0.6 for output. Peak-hour prices will be double the off-peak rates. Please plan your usage accordingly.
------------------------------------------------------- Hoje em sites como openrouter o valor é de $0.16 output .
Input cache hits (per 1m tokens) - $0.003 Vs. $0.022 Vs. $0.007
Input cache miss (per 1m tokens) - $0.15 Vs. $0.66 Vs. $0.22
Output (per 1m tokens) - $0.6 Vs. $1.98 Vs. $0.66
This is taken from https://api-docs.deepseek.com/quick_start/pricing, and it's comparing only off-peak hours pricing. It looks like V4.1 Flash is cheaper than the current 0731 flash model, and much cheaper than the current V4 Pro model.
0.007 -> 0.003
0.22 -> 0.15
0.66 -> 0.60
Each one is now cheaper.
[0]: https://api-docs.deepseek.com/quick_start/pricing/
As a consumer I feel like hansel and gretel combined, deepseek could be the witch.
— hmm — 0x2D696370 — little-endian bytes: 70 63 69 2D = 'p','c','i','-' — hmm — WAIT — WAIT — !!!!! — *WAIT — WAIT — WAIT — WAIT — WAIT — WAIT — *HOLD ON — HOLD ON — HOLD ON — WAIT — WAIT — WAIT — WAIT — WAIT — WAIT — WAIT — !!!!!!!! — *WAIT — WAIT — WAIT — WAIT — WAIT — WAIT — WAIT — WAIT — *OK — WAIT — I THINK I FINALLY SEE THE WHOLE PICTURE — I NEVER READ IT — AND — THE LAYOUT — hmm — !!!!! — *WAIT — WAIT — WAIT — WAIT — WAIT — WAIT — WAIT — WAIT — WAIT — HOLD ON — HOLD ON — HOLD ON — HOLD ON
Then gave the same to Sonnet 5 and it was done 15 - 30 minutes later. I tried v4 pro both in claude code and codewhale with similar results. Haven't tried the new deepseek harness.
It built this whole IaC plugin from scratch: https://github.com/fllstck/nebius-alchemy
Wow. Imagine OpenAI/Google/Anthropic doing this! Nope.