DeepSeek V4.1 Flash hands-on tests and comparisons

In an AI news roundup, DeepSeek-V4.1-Flash-1017 at max setting scores 57% on Artificial Analysis' defense-only Cyber Index, ranking first among publicly available models.

Using OpenCode Go list prices, DeepSeek V4.1 Flash peak output costs $1.2 per million tokens versus $10 for Kimi K3.1 Pro, an 8x gap; the $80 vs $15 quotas work out to roughly 66M vs 1.5M output tokens, a 40x difference.

The author argues Terminal-Bench scores are gamed: on Agent Arena blind tests Kimi K3.1 Pro ranks 11th (+7.6%) versus DeepSeek V4.1 Flash 1017 at 13th (+4.4%), while costing about 5x more.

A reply notes that running V4.1 Flash locally needs 8x H200 or a community 4x RTX PRO 6000 Blackwell setup, and hopes consumer GPUs can handle it soon.

Terminal-Bench 4.0 results on DSH put V4.1 Flash-1017 around GLM-5.3-Flash and Haiku-5.5, above Opus 4.8, at overwhelmingly lower cost, with high reasoning effort recommended.

A local test of DeepSeek V4.1 Flash 8K with Strata on Windows 10, an RTX 5090 32GB and 64GB DDR5 measured about 8–10 tok/s at 31.4GB VRAM, with commenters arguing API use is cheaper than the power bill.

Hands-on: DeepSeek V4.1-Flash runs via Strata on a 4070 SUPER 12GB with 64GB DDR4, despite the setup warning that only RTX 30/50 series are supported.

DeepInfer runs on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, defaults to the 1M-context V4.1-Flash-1017, and offers deterministic sampling and structured outputs.

China's Supercomputing Internet promo: new users get 10M free tokens, and ¥0.1 buys a 50M DeepSeek-V4.1-Flash-1017 token pack (list ¥20.5), daily at 10am until Nov 10.

At 131K context vLLM's DEP32 stays within 6% of DeepInfer's throughput, while a GB300 NVL72 with P/D and Wide-EP reaches 954K tokens/s per GPU and 826 tok/s single-user decode, with launch commands included.

A user comparing with Opus 5.5 says DeepSeek V4.1 Flash 1017 has the best intelligence-per-parameter among self-hostable models and handles almost everything, making a Claude subscription near diminishing returns.

A user ran the IQ2_XS GGUF quant of DeepSeek-V4.1-Flash-1017 in Strata on a 5070 Ti with 96GB DDR5 and got 10.7 tok/s.