DeepSeek V4.1 Flash hands-on tests and comparisons

A 24-hour AI roundup reports DeepSeek released its in-house inference software DeepInfer for free (155 trillion tokens per day per cluster, Blackwell and Ascend 950 support) and confirms V4.1-Flash-1017 remains the current model.

The user says DeepSeek V4.1 Flash is demanding to self-host: reaching 1M context takes four 128GB unified-memory DGX Sparks, with a side note on security concerns.

Citing Vals AI's cyber range test, a user notes DeepSeek V4.1 Flash scores 16.3/21 versus GPT-6 Sol's 15.7/21, while being open-weight and cheaper.

A user relays a community evaluation putting deepseek-v4.1-flash-1017 roughly on par with GPT-5.6 Sol xhigh, noting it was run inside DeepSeek Harness.

The piece compares China and US token totals, output volume and revenue per token, notes DeepSeek's $0.60 vs Opus's $25 output price, and estimates 1.5M Ascend 950s on DeepInfer could produce ~190T output tokens per day.

Rapid-MLX 0.16.7 speeds up its GPTQ v2 quant of DeepSeek-V4.1-Flash-1017 to 33–45 tok/s single-stream on an M3 Ultra via lighter drafting, fused mixed-bit MoE and a once-per-verify vocab head.

A reply argues DeepSeek's fused attention-MLP block lets Flash-1017 cut latency while scaling compute on a larger token corpus, contrasted with another model's retrieval-augmented decoder.

A user who has GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017 running locally will spend the next week trying Zach's local-codex-proxy to connect them.

DeepSeek released DeepInfer under Apache-2.0, saying it powers high-throughput, low-latency inference for V4.1 Flash, but reports say it only runs that model on Nvidia SM100 or Ascend 950 and covers offline inference only, without KV cache or prefill/decode separation.

A user asks which tool suits DeepSeek V4.1 Flash 1017 best — Pi, DSH or Claude Code — without sharing any test results.

A team uses uncensored DeepSeek V4.1 Flash to run 5–10 subagents pentesting its own app, then has GPT-6 Astra confirm and patch the findings.

A commentary says Chinese Flash models that lose to DeepSeek V4.1 Flash-1017 on both price and performance should retrain, and flagships behind Claude Fable 5.1 need not rush to ship.