DeepSeek V4.1 Flash tops independent open-weight evaluations

• DeepSeek's new Flash is a new model family: kimmonismus's launch rundown describes a 552B MoE with a causal encoder-decoder split that activates 8B parameters for input and 16B for output.
• The serving design cuts KV-cache HBM use to one quarter of the prior generation, while stochasticchasm's FP4 cache note calls out the combination of FP4 cache and compressed sparse attention.
• Independent scorecards put it near the top of open-weight comparisons: ValsAI's index update names it the new Vals Index leader, while ArtificialAnlys' scorecard gives it a 40 Intelligence Index score at $0.27 per task.
• API model IDs are changing underneath users: aibuilderclub_'s routing summary says V4-Pro requests will route to V4.1 Flash at Flash prices until V4.1-Pro ships.

DeepSeek's model card gives decoder layers a projection of final encoder states rather than independent global KV histories. It also exposes reasoning_effort as an integer from 1 to 100, while the official release redirects legacy Flash aliases to the new model.

The 40-layer stack is split into 20 causal encoder layers and 20 decoder layers. During decode, each decoder layer gets its global KV state from a projection of the encoder's final hidden states; it still computes its local sliding-window attention from current-layer states.

That arrangement separates input prefill from output generation and is the mechanical basis for the 8B and 16B active-parameter counts. nrehiew_'s CED breakdown

CED removes repeated global KV construction, while Compressed Sparse Attention 2, or CSA2, reduces which context state each layer carries and searches. The model card assigns every attention layer one of three static reuse modes:

• Full: performs a fresh selection without reuse.
• Reindex: reuses an earlier layer's main KV and indexer K, then performs selection again.
• Reuse: carries forward the earlier selection output.

A hierarchical sparse indexer confines later layers to a candidate pool selected by the first Full-mode layer. FP4 main-KV caching, with quantization-aware training, brings global cache state to 890 bytes per token, roughly one quarter of V4 Flash; DeepSeek reports one-quarter HBM and one-eighth SSD requirements versus the preceding generation in its release post.

The independent leaderboards measure different workloads and produce a more useful picture than a single launch table.

• Vals: V4.1 Flash rose 21 places and 19 points over V4 Flash on SkillsBench, according to ValsAI's SkillsBench post. It also took the open-weight Code Migration lead at under $1 per test, compared with $24.91 for second-place GLM 5.3 in ValsAI's Code Migration post.
• Code Arena: the initial AutoEval result moved from 1,582 for V4 Flash High to 1,620 for V4.1 Flash, a +38-point gain; arena's early Code Arena result notes that a reward model cast those early votes and live human voting remains in progress.
• Artificial Analysis: its independent model page places V4.1 Flash sixth of 113 models at 40 on the Intelligence Index, with 190.1 output tokens per second.
• DeepSeek's own table: DeepSWE v1.1 rises from 54.4% for V4 Flash to 74.2%, a +19.8-point change, in kimmonismus's DeepSWE comparison.

Artificial Analysis puts V4.1 Flash's cost at $0.27 per Intelligence Index task, but its evaluation consumed 250M output tokens, against a 130M-token median. Its separate scorecard reports 89,000 tokens per task, 62% above V4-Pro's 55,000.

One practical run points in the same direction: bridgemindai's lava-lamp test measured 344 tokens per second but spent 23.5M tokens and completed later than Fable 5.1 and GPT-6 Astra.

The model card's 1-to-100 reasoning control directly adjusts the length penalty during training and inference. DeepSeek also shipped experimental Team Mode in its Harness, as the Team Mode announcement put it.

The post-training recipe explicitly rewards task completion, delegation and inter-agent communication, then penalizes inefficient coordination, according to nrehiew_'s Team Mode analysis. In an early Nacre-fan trial, teortaxesTex's agent-team test found that a team spent heavily on unnecessary optics research and interfered with itself, while a solo run was cheaper and produced the preferred result.

DeepSeek's first-party API now uses deepseek-flash; its release notice says deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1. From September 14, deepseek-v4-pro follows the same route until V4.1-Pro arrives.

Day-one infrastructure support arrived across several surfaces:

• woosuk_k's day-zero vLLM note announced vLLM support.
• cline's Cline instructions lists V4.1 Flash in Cline's model selector.
• Teknium's Hermes Agent announcement says it is live through Nous Portal and other hosts.
• ollama's cloud rollout began with Ollama Max and Team accounts before wider subscriber rollout.

Baseten's baseten's day-zero API post adds a US-only, zero-data-retention endpoint with a 1M-token context window.

The weights and repository carry an MIT license, and the model card includes local-inference conversion code plus the deepseek-recipe prompt-encoding libraries. Alongside the 552B backbone, it carries 196B Engram conditional-memory parameters accessed through sparse token lookups.

DeepSeek's release calls for operators with 2,000 GPUs and a storage cluster to discuss large-scale deployment. That framing matches an early teortaxesTex's weight-size estimate of roughly 500GB on disk and a Hugging Face community thread asking for a smaller Flash Lite variant.