DeepSeek V4.1 Flash hands-on tests and comparisons

An overview of the October 17 V4.1 Flash '1017' release: MIT-licensed open weights, 1M-token context, and a vendor-reported DeepSWE coding score of 77.0, runnable on private servers.

A quoted comparison says DeepSeek V4.1 Flash 1017 found 54 vulnerabilities for $7.53, seven more findings at roughly 16.6x lower cost than the other model.

On a single RTX 5090 with the Maya-S quant, DeepSeek Harness built a voxel sky world from one prompt in 74.1 min, scoring 8.5/10 in browser checks at 22.9 tok/s code decode, with blocks breaking up close.

On the same procedural Three.js coastal diorama prompt, V4.1 Flash took about 1 hour and 78M tokens versus MiniMax 3.1 Flash's ~2h17m and 60.3M tokens, both producing lighthouse, boat, water shaders, birds and night mode.

The author runs a single-agent MoA setup combining DeepSeek V4.1 Flash with Gyro S, letting Gyro serve as the perception layer, hoping it finishes before morning.

A user reports running DeepSeek V4.1 Flash at Max setting almost nonstop for two days in OpenCode Go and consuming only 11% of the monthly quota.

The poster shows DeepSeek-V4.1-Flash-1017 generating multiple simulations in one shot via direct API calls, with no harness, tools, or external assets, as a raw-capability demo.

A 748B model in Maya-S quant ran on a single 5090 via RAM/NVMe offload, finishing a voxel-world build in 74.1 min with 100,575 tokens written, 8.5/10 browser check and 22.9 tok/s code decode.

The author calculates $20 buys ~2.9B tokens on Claude Pro vs ~1.8B on DeepSeek V4.1 Flash because ~96% of billing is cache, yet at API list price those tokens cost ~$1,650 vs ~$33.

The author estimates ~35 tok/s and ~1.1B tokens/year locally on a 256GB M5 Ultra, costing $394 in electricity versus ~$663/year for equivalent off-peak API output, concluding the laptop is overspend if bought only for local LLMs.

A user complains V4.1 Flash got slower, yet measured throughput is still 150 tok/s on average and up to 169 tok/s, which the poster frames as rising expectations.

A tester lists 1017 agent flaws: budget counting inside thinking breaks, 'check status first' prompts fail, ambiguity handling is poor, and constraining it triggers a power-saving mode; coding remains strong, hitting 2K 60fps on a 3080Ti.