DeepSeek V4 Flash on my architect bench: 0.717 and ₽0.022 per solved task
DeepSeek V4 Flash scored 0.717 on my ЛИИ-Архитектор-Bench-RU: 41 of 50 tasks solved, TTFT 4.72s, 23.7k reasoning tokens, ₽0.022 per solved task.
My intuition was ready to write DeepSeek off after V4-Pro: it landed on the board at 0.698 and never caught the top three. The measurement corrected me. Flash is still below GLM-5.2, Qwen 3.6 and Gemma 4 — but it beat V4-Pro at 4.8× lower price and is now the cheapest ₽/solved on the entire board.
Same bench as before: 50 real tasks mined from my own working sessions. Not HumanEval, not "guess the right import" — my actual workload: strategic synthesis, architecture decisions, documents, research distillation, some code and ops. Scoring is a three-judge ensemble: Gemini + GPT-5.1 + Opus 4.8. A single judge is noisy on tasks like these, so I look at the average.
Today's board
| Model | Score | Note |
|---|---|---|
| GLM-5.2 | 0.809 | the quality crown, "send and wait" mode |
| Qwen 3.6 | 0.779 | strong daily line |
| Gemma 4 | 0.772 | the practical self-host pick from the last run |
| DeepSeek V4 Flash | 0.717 | 41/50, TTFT 4.72s, 23.7k reasoning, ₽0.022/solved |
| DeepSeek V4-Pro | 0.698 | the previous DeepSeek baseline |
| Leading RU cloud LLM* | 0.477 | the sovereign incumbent lost to open models |
Artificial Analysis puts Flash at Intelligence Index 45. That's roughly Haiku 4.5 territory (43), below Sonnet 5 (53). My short version: "last-generation Sonnet class at 1/50th the price." Not a Sonnet 5 replacement everywhere. But for a stream of architect tasks where price per solved task matters, that's a different shelf.
The real twist is memory
DeepSeek V4 Flash is a 284B-A13B MoE, MIT license, 1M context, $0.09/$0.18 per Mtok. On paper that's still a big machine: you hold the whole checkpoint in memory, not the 13B active. In the last sports run this is exactly why I filed DeepSeek under "cluster budget."
The twist is the weight format. The native mixed FP4+FP8 checkpoint is 148.7 GiB. That fits 2×RTX PRO 6000 — two 96GB cards, 192GB total. Pure FP8 is 273.9 GiB and doesn't. The difference between "buy two workstation cards" and "build a big box" came down to the checkpoint format.
Community reports on exactly this pair of cards show 100–220 tok/s single-stream and around 10K tok/s prefill. Not my number yet — I haven't run it on my own box. But the order of magnitude is enough to stop treating Flash as an API-only model.
And a practical detail: it's already hosted on Yandex AI Studio. You can poke it inside the Russian contour before buying any hardware. For a trusted-AI business that's the right order: verify the discipline on your own tasks first, think about capex second.
The honest caveat
Flash "thinks" long: TTFT 4.72s and 23.7k reasoning tokens over 50 tasks — nothing like Gemma's instant 1.24s with zero reasoning. SM120 serving for MoE this size is still fork-heavy. And this is my bench on my tasks. It doesn't prove Flash is stronger for your product.
It proves something else: on my architect workload, DeepSeek V4 Flash became the cheapest model that actually solves tasks on the board — and for the first time looks like a real candidate for 2×RTX PRO 6000 instead of a pretty row in someone else's API.
Full board, methodology and raw runs: bench.csylabs.com/architect.