DeepSeek V4.1 Flash hands-on tests and comparisons

On NexBench, which measures discovering and executing real exploits plus price-to-performance and token efficiency, DeepSeek-V4.1-Flash-1017 scored highest of 18 models.

On VulcanBench's Frontier v4 suite, V4.1 Flash lands in the 80–90 good-enough range with far lower cost per task than GPT-6.1 Sol and Opus 5.5, though raising effort barely helps.

One try each on the same prompt: cloud Opus 5.5 finished in 33.2 min for about $7.82, while V4.1 Flash on four DGX Sparks took 49.7 min at no API cost, both scoring 9.0/10 in a browser check.

V4.1 Flash uses 890 bytes of KV per token, a quarter of V4 Flash; at an 85% hit rate the cache avoids 3.4–4.1x more GPUs, and 100T tokens of context takes 89 PB.

Suggests four people chip in 6k each for four DGX Spark/GB10 units to run an uncensored V4.1 Flash locally for unlimited free tokens, with no benchmarks.

The Apache 2.0 release bundles DeepSeek's FlashMLA, DeepGEMM and DeepEP kernels with DSpark speculative decoding and batch-invariant inference, running on SM100 GPUs and Ascend 950 NPUs.

The dsh-our-free-model plugin at 38.6K stars adds DeepSeek V4.1 Flash, Kimi K3 and 14 more models to dsh's picker with no login, API key or usage cap.

Official summary: stronger benchmarks, faster and shorter outputs, live on web, app and API, plus the production inference stack DeepInfer open-sourced to cut community cost and energy.

Maya-XS compresses the 748B model to about 120.6GB with ~1.6-bit quantization and tiers experts across VRAM, RAM and SSD, lifting speed from 17 to 65 tok/s, with three V100s reaching 71.7 tok/s at 68K context.

A Pi 1.3.0 multi-agent pipeline using V4.1 Flash handled research, review, coding and merge gates, catching one flaky test, and the final npm test run passed 304/304 on a fresh clone.

The author chains Telegram, Hermes Agent and the local magpie gateway to use WorkBuddy's limited-time free DeepSeek V4.1 Flash, lists four setup commands, and reports zero quota used after two days of long tasks with one-command model switching.

The experimental Maya-XS build keeps all 384 experts in 120.6GB of RAM, matches the original's next token 87.5% of the time with 8.7% higher loss, and needs about 280GB to download.