DeepSeek V4.1 Flash hands-on tests and comparisons

Local deployment test: DeepSeek V4.1 Flash runs on dual DGX Spark 256GB, a 256GB Mac via oMLX and four DGX Sparks with TP4 for browser world-building agents, no cloud API.

Third-party platform Kenari says the free DeepSeek V4.1 Flash and GLM 5.3 Flash tiers end tonight at 13:00 UTC and move into plans starting at Rp49k/month.

The poster notes several inference providers are using the open-source deepinfer to speed up and cut prices for DeepSeek V4.1 Flash, without specific numbers.

A Japanese user wants to try DeepSeek V4.1 Flash 1017 but believes local deployment needs at least 512GB of memory.

A line-by-line review of antirez's pure-C DwarfStar finds it supports V4.1 Flash among few models, that 126 t/s is 16-session aggregate throughput, and Macs need 96GB+ RAM.

A user's chart shows V4.1 Flash-1017 modestly improving on ARC-AGI-2 and slightly beating Gemini 3.1 Pro Preview, with a side complaint about missing token-axis options.

Since the HF card omits Cyber/ExploitGYM results, the author benchmarks V4.1 Flash 1017 against GLM-5.3, doubting DeepSeek beats GLM.

On the new FrontierCode v1.1 benchmark, Codex with DeepSeek V4.1 Flash scores 52.6%, below Claude Opus 5.5 at 56.0% and above GPT-6.1 Sol and GPT-6 Astra.

An RSNA competitor used GPT-6.1 Sol labels to train six models reaching about 0.941 gold score, and says DeepSeek V4.1-Flash-1017 performed poorly on the same task.

A user notes DeepSeek open-sourced DeepInfer, the inference engine behind V4.1 Flash, with Huawei Ascend support, and asks whether people read such engine code.

The author used deepseek-v4.1-flash-1017 to one-shot an explainer video built entirely from CSS, SVG and HTML, with no external assets, harness or tool calls.

A detailed breakdown of running the 763B V4.1 Flash checkpoint (510.3GB, 6 of 384 experts active per token) on 3090/4090/3060 builds via expert tiering, noting Engram width matters more than expert bits.