DeepSeek V4.1 Flash hands-on tests and comparisons

Roundup of the week's open-weights news, including DeepSeek's DeepInfer release and V4.1 Flash building a 3D castle in 48 minutes for $0 versus Sonnet 5.5's 15 minutes and $1.2.

Five models were asked to build and render a 15-second Three.js flyover of Riyadh's KAFD from OpenStreetMap data; DeepSeek V4.1 Flash, run on 4 DGX Sparks, placed fourth behind Sonnet, MiMo and Haiku, ahead of Qwen.

The author recalls only hoping to run Qwen3.6-27B on a used RTX 3090 in June, and now runs Qwen3.8-Flash-Next, GLM-5.3-Flash and, barely for now, DeepSeek V4.1 Flash 1017.

A four-step setup adds a plugin to DeepSeek's dsh agent harness so V4.1 Flash, GLM 5.3 Flash and Kimi K3 appear in the model picker with no login or subscription.

The author pushes back on claims that DeepSeek isn't truly open source for withholding training data and code, noting DeepSeek V4.1 F-1017 scores 45 on the AA intelligence index versus 11 for K2 Think V2, which does release them.

A user says DeepSeek V4.1 Flash beat GPT 6 Sol on a Vals AI multi-step offensive-security test and can run locally, mocking OpenAI's access gate.

A post says OpenCode lists nine models free for a limited time, including DeepSeek V4.1 Flash (paid price $0.30 in / $1.20 out per 1M tokens), GLM-5.3 Flash and MiniMax M3.1 Flash, with no end date published.

A user's local-workflow stack claims DeepSeek V4.1 1014 handles about 98% of what Claude does, plus 20% Claude won't do, at roughly one-thirtieth the price.

The 1.61 bpw EXL3 quant fits one DGX Spark at ~112.9 GiB with 94.68% assistant-token top-1 agreement; the 2.50 bpw two-Spark version reaches 96.71%.

On the same prompt MiniMax leaned cinematic while DeepSeek emphasized environment and detail; pricing is listed at up to $0.30/$1.20 per million tokens at peak.

A user runs DeepSeek V4.1 Flash 1017 locally with EXL3 2.95bpw on two 3090s without NVLink, reaching 783 t/s prefill and planning to share the setup.

RouterLink lists DeepSeek-V4.1-Flash-1017 with shorter outputs at the same API price, DeepSWE v1.1 up from 74.2 to 77.0, and a 1M-token multimodal MoE.