DeepSeek v4 Pro Review (Real World Tests)

DeepSeek V4 Pro's killer feature is what happens between the code blocks

DeepSeek V4 Pro and V4 Flash shipped this week. The marquee spec is a hybrid attention architecture that compresses irrelevant context: at a 1M token context window it uses 27% of the compute and 10% of the memory of the prior model. Cherry-picked benchmarks the creator flagged: 93.5% LiveCodeBench, 80.6% SWE-Bench, and a ~20-point lead on Simple QA over other open models. Cost on pay-as-you-go API was less than $0.30 for 150 requests (102 on V4 Pro in Kilo Code, 48 on V4 Flash in Hermes). No coding plan exists — only API key, with three modes: non-think, think high, think max.

Tool-call chain-of-thought is the differentiator

On a Chinese 3D building prompt, Kimi 2.6 and MiMo 2.5 Pro ship code quickly with shallow reasoning. V4 Pro generates a ~1000-line code block but interleave heavy self-critique — the creator highlights lines where the model writes "actually a cleaner approach" and "let me also think about the x-ray mode". It catches type errors mid-build on the crypto dashboard task and only then commits the fix, instead of dumping a broken file. The trade-off is speed: this project took 40–45 minutes longer than Kimi or MiMo.

Raw output, not so much

Ranked across the same one-shot suite the channel uses for every Chinese model: - 3D building: Kimi 1st (cleanest UI), V4 Pro 2nd (most features, messy labels), MiMo last. - Space shooter: Kimi 1st (gameplay works), V4 Pro 2nd (graphics clean, can't kill enemies), MiMo last. - Minecraft/voxel world: Kimi 1st (laggy at 7 FPS but runs), MiMo 2nd (broken world, smooth), V4 Pro last (black screen). - Crypto dashboard: V4 Pro blanked on VPS, returned a localhost link, recovered after a prompt fix.

The creator's read: it's "a prompt issue, not a model issue" — vague one-shots underfeed V4 Pro's need for explicit specs.

Use it as an agent, not an IDE model

A side-by-side V1 vs V2 space shooter in Hermes showed V4 Pro's strength: V1 was a single file write, V2 used the same long chain-of-thought to add an overheat mechanic and working fire. Kilo Code also currently can't toggle reasoning modes for V4 Pro — base URL stays api.deepseek.com (no /v1).