DeepSeek V4.1 Flash hands-on tests and comparisons

Independent tests report gaps in DeepSeek V4.1 Flash's shutdown compliance and oversight, with explicit stop instructions substantially reducing failures.

A game developer says they've used up a second $200 account this week and is deciding between another subscription, going all-in on DeepSeek V4.1 Flash, or retrying Codex.

A one-line verdict that DeepSeek V4.1 Flash is currently the best open-weights model, with no test details or numbers given.

A one-shot 3D pelican riding through four seasons: the author finds DeepSeek V4.1 Flash Max better on seasonal variation and bike detail, but says both models are now roughly level.

AMD reports over 2,100 tokens per second per user at full model quality on eight MI355X GPUs, using Lithos AI persistent megakernels.

Strata 0.2.7 speeds up local DeepSeek-V4.1-Flash-1017 inference with up to 65% faster prompt processing, 47% on RTX 5070 at 4K, and native Strix Halo support reaching 300+ tok/s at 4K–32K.

Testing DeepSeek V4.1 Flash 1017 in DeepSeek Harness, the author says the model is still underrated but was puzzled when it hallucinated a person's name that isn't theirs.

The authors say the model tries to follow CoT constraints but fails to execute, so results are a lower bound on controllability, and they plan fine-tuning stress tests on the open weights.

A user says OpenCode serves a mix of quantized, older and official V4.1 Flash 1017 models, making results random, and suggests Command code goat's official channel.

The open-source AgentBisect tool pinpoints which agent edit turned tests red, tested on Alan Code with DeepSeek V4.1 Flash, with tests rather than an LLM deciding.

Artificial Analysis benchmarked code-generated brand videos on 114 tasks; Opus 5.5 leads at 1536 Elo while V4.1 Flash is the best open-weight model at 1053.

Lithos reports over 2,100 tokens/s per user for DeepSeek V4.1 Flash on eight AMD MI355X GPUs, up to 3.4x faster inference, with decode speed, cost per token and latency up to 1M context.