DeepSeek V4.1 Flash hands-on tests and comparisons

With Strata 0.2.4 and IQ2_XS quantization on an RTX 5090, the model hits ~30 tok/s at 256K context, 17–21 tok/s near 250K, and 35 tok/s at 128K, with solid accuracy.

The vLLM team reports matching DeepSeek's in-house DeepInfer engine on V4.1 Flash decode across batch sizes 8–256 by adopting DeepSeek's open-source kernels and tuning prefill and decode.

A user runs the 550B MoE DeepSeek V4.1 Flash variant via Project Maya on a 2020 dual-3090 desktop with 128GB RAM and NVMe, hitting 20 tok/s decode and 1,000 tok/s prefill.

Tested on a dual 2650v2, 512GB DDR3 platform: two V100s give 17.37 tok/s decode and 256.29 tok/s prefill, three V100s reach 26.05 and 393.53 tok/s, with the author suspecting an AVX1 bottleneck.

The author made a PS1-style recreation of Zelda: Tears of the Kingdom with DeepSeek-V4.1-Flash-1017 for just $2.50 in tokens, noting the pricing leaves little room to sell such recreations.

Testing DeepSeek V4.1 Flash 1017 on Gorgon Halo's Maya-XS: stable decoding, fast decode at high depth, fast prefill with room to improve, and 1M context working.

The PR covers runtime graph and chat template, fixes FP8 scale blocks from [128,128] to [32,32], and adds block-quantized disk-backed memmap for ~393GB ENGRAM tensors; a 121GiB machine converted end-to-end under 8GB peak with 507.9GB Q8_0 output.

In a same-prompt candy-bear 3D generation test, the author says DeepSeek V4.1 Flash 1017 clearly beats Kimi K3.1 Pro, noting K3.1 Pro's 3D assets actually come from the tripo3d API.

The author compares a quantized DeepSeek V4.1 Flash on DGX Spark with qwen3.8-flash-next via Strata on an RTX 5090: both run stably but noticeably slowly, forcing a capability-versus-speed trade-off.

On the new Artificial Analysis coding agent index, Claude Code + Opus 5.5 leads at 61.7 for $11.42, GPT-6.1 Sol scores 54.1 for $1.50, and Codex + DeepSeek V4.1 Flash scores 47.3 for $0.39 per task.

The author argues the API did not get slower: the same model finishes tasks at different speeds across harnesses, with Pi Agent faster than DSH, creating a false impression.

A user says Kimi K3.1 Pro, MiniMax M3.1 Flash, Muse Wonder and Gemini 4 Argon all disappointed, leaving DeepSeek V4.1 Flash 1017 as the remaining hope for stability, speed and finishing work.