DeepSeek V4.1 Flash hands-on tests and comparisons

llama.cpp build b11870 adds deepseek41 architecture support for DeepSeek-V4.1-Flash, converting the Engram table to BF16, gathering Engram rows in-graph and computing compressed KV layers once, enabling GGUF local runs.

A user reports that running DeepSeek V4.1 Flash locally overflows VRAM plus DRAM and falls below 20 tok/s.

On the 290-task index v2.0 cost frontier, Codex with DeepSeek V4.1 Flash scores 47.3 at $0.39 per task, versus 54.1/$1.50 for GPT-6.1 Sol, 54.7/$7.04 for GPT-6 Astra and 61.7/$11.42 for Claude Code with Opus 5.5.

Fireworks claims over 1,521 tokens per second and shows Tetris code generated in 11 seconds; the fast tier is waitlisted and the figures are demo conditions, with real-world verification still needed.

A news roundup notes DeepSeek-V4.1-Flash-1017's Hugging Face downloads rose from 27.7K to 73.2K in four days with 2.77K likes, as DeepInfer and community quantizations close the ecosystem loop.

Given the same prompt for an isometric miniature city block, DeepSeek V4.1 Flash took 27 minutes and $0.90 while MiniMax M3.1 Flash took 30 minutes and $0.49; the author calls MiniMax the better value.

PR 31086 landed in master with a new deepseek41 architecture (~2,800 lines across 39 files), its own conversion script, --dspark packing MTP blocks into a DFlash drafter, and a vision tower; no tagged release yet, so build from master.

A cost chart shows DeepSeek V4.1 Flash 1017 at $0.3 input / $1.2 output per unit, roughly one-sixth of Kimi K3.1 Pro's $2 / $10.

At the same DeepSWE 1.1 coding score of 0.742, DeepSeek V4.1 Flash costs $0.22/$0.66 per M tokens versus Opus 5.5's $4/$20 and GPT-6 Astra's $10/$50.

A weekly open-ecosystem roundup says DeepSeek-V4.1-Flash beats GPT-5.6 Sol on Terminal-Bench 4.0 at 20x lower cost, alongside LFM2.5 on-device agents and Coinbase's Qwen fine-tune.

A user shares a price comparison showing Kimi K3.1 Pro priced against Claude Sonnet 5, arguing it is too expensive and that DeepSeek V4.1 Flash 1017's value keeps rising by contrast.

Jarvislabs Token Factory now serves DeepSeek V4.1 Flash (1017) with a 1M context at $0.07 input, $0.003 cached input and $0.49 output per 1M tokens, 180-200 tok/s, compatible with OpenAI SDK, Claude Code, Codex and OpenCode.