DeepSeek V4.1 Flash hands-on tests and comparisons

The author finds DeepSeek Flash's quality-per-cost unremarkable and says the 1017 version still over-thinks, hurting token efficiency, yet its official API is good enough and far cheaper than OpenAI or Anthropic at list price; they also criticize the update for optimizing for demos over practical backend work.

A user runs an abliterated V4.1 Flash 1017 on a Mac Studio via DwarfStar and offline Hermes, patching 62 tensors in a 467GB APFS clone and calling its way of seeing things as unfamiliar as R1's.

Recap of the official release: V4.1-Flash-1017 ships on web, app and API with MIT-licensed weights, and Terminal-Bench 4.0 rises from 31.2 to 42.3 per vendor numbers, not independently retested.

DeepSeek released DeepInfer, the in-house inference framework behind large-scale DeepSeek-V4.1-Flash serving, plus the production token-conversion component deepseek-recipe, so others can reproduce official API behavior in their own environments.

On identical tests across two DGX Sparks, DeepSeek scored 93 on tool calling versus GLM's best 96 (but GLM finished 8 minutes faster); at 384K context DeepSeek EXL3 2.9bpw got 48/48 while GLM quantizations scored 35–48, and at 480K DeepSeek held 20/20 versus GLM fp8's 14/25.

The post says V4.1 Flash costs roughly one-thirtieth of Claude Opus 5.5, framing the shift as cheaper everyday and sub-agent work rather than raw performance.

The author connects free DeepSeek V4.1 Flash to Codex via Magpie for real work, with GPT-6 Astra orchestrating to split tasks and ease token limits.

A rundown of what DeepSeek has opened around V4.1 Flash: MIT weights, FlashMLA, DeepEP, DeepGEMM, 3FS, DeepInfer and the deployment recipe, with only training data kept closed.

Cloudflare's changelog announces DeepSeek V4.1 Flash 1017 is available on Workers AI with a 1M token context window and built-in reasoning.

The post cites specifics: MIT open weights around 510 GB, a 552B backbone plus 196B Engram activating only 8B (prefill) / 16B (decode) per token, API at $0.15/M input and $0.60/M output off-peak, with quantized builds running on high-RAM machines or a few GPUs.

A Zhihu technical report on DeepInfer, the inference stack behind V4.1 Flash, covers prefill/decode hardware choices and KV-cache targets plus real cluster data: 155 trillion tokens per day and $38 revenue per GPU-hour at off-peak pricing.

Recomputing 305,287 real agent requests at list rates, the author found DeepSeek V4.1 Flash 1017 costs ¥0.1995 per million tokens, below GPT-6 Luna's ¥0.3227 and GLM-5.3-Flash's ¥0.3634, contradicting the Artificial Analysis ranking.