DeepSeek V4.1 Flash hands-on tests and comparisons

A user asks whether a ready-made setup exists to run DeepSeek V4.1 Flash 1017 as subagents in Claude Code or Codex while keeping Opus/Sol as the main orchestrator.

On AgentSky, switching Codex's model field to deepseek-flash produced a landing page with DeepSeek V4.1 Flash in 4 minutes 48 seconds at $0.08 metered cost.

A user asked V4.1-Flash-1017 to write a song about itself and build a JavaScript animation; it took 20M tokens and $0.30, with the result shown in a video.

A ~114GB VQ build served DeepSeek V4.1 Flash 1017 on one DGX Spark at 39.4 tok/s decode and ~1,145 tok/s prefill with a 110K prompt; Artificial Analysis rose 39 to 45 and cost per task fell $0.49 to $0.26.

Rapid-MLX shipped an MLX build of DeepSeek-V4.1-Flash-1017 keeping all 40×384 experts with mixed 2/3/4-bit quantization, KL divergence 0.439, and ~20 tok/s at 206.5GB resident on an M3 Ultra.

The author rebuilt a 2.70-bit quant on the 1017 weights, getting 141.3 tok/s structured output, 100.7 code and 59.2 prose on two DGX Sparks, with KLD 0.0243 vs native 1017, 95.47% top-1 agreement, and HumanEval+/MBPP+/IFEval within noise.

Release summary: MIT-licensed open weights, 1M context, DeepSWE coding score up from 74.2 to 77.0, and a search API at $0.75 per thousand calls.

The author admits the earlier test used an old checkpoint and re-runs 1017 vs the previous V4.1 Flash: 1017 delivers better choreography, visual effects and overall feel at the same price and completion time.

Design Arena projects DeepSeek-V4.1-Flash-1017 to rank #3 among open-weight models on its website leaderboard, a 16-place jump over V4-Flash-0731, behind only Kimi K3 and MiMo-V2.6-Pro.

With SGLang on 4-way tensor parallelism, prefill holds 3,500–4,000 tok/s from 4K to 274K input; code decodes at 83 tok/s, prose ~45 tok/s, and 274K-context decode ~47 tok/s with ~77 s TTFT.

The author ran Opus 5.5 and DeepSeek V4.1 Flash overnight to batch-generate low-poly 3D objects, harvesting 1,080 results in the morning.

Running V4.1 Flash 1017 locally in Maya on four 3090s with 256k context gives about 35-45 tok/s generation and 500-1000 t/s prefill; the model is nearly 400GB.