DeepSeek releases 552B-parameter V4.1 Flash multimodal model

• V4.1 Flash is a 552B-parameter multimodal MoE with 8B active parameters on input and 16B on output, a design that a LocalLLaMA launch post reproduces from DeepSeek's release.
• DeepSeek reports large agentic gains over V4 Flash, including 82.7% to 90.6% on Terminal-Bench 2.1 and 37.7% to 54.8% on AutomationBench, results that teortaxesTex's benchmark post summarized from the official chart.
• The new Causal Encoder-Decoder design cuts global KV cache to 890 bytes per token, about one quarter of V4 Flash, while the LocalLLaMA launch post says HBM demand falls to a quarter and SSD demand to an eighth.
• V4 Pro API traffic will temporarily run on Flash at Flash prices until V4.1 Pro ships, as aibuilderclub_'s rollout post reported.

The model card ships a reference prompt encoder with tests for interleaved images, tool calls, and mid-conversation system messages. In bridgemindai's BridgeBench lava-lamp run, V4.1 Flash generated 23.5M tokens at 344 tokens per second for $0.33.

• A 552B backbone MoE, native image and text input, and autoregressive text output, according to DeepSeek's official release announcement.
• A Causal Encoder-Decoder architecture with a 20-layer causal encoder and 20-layer decoder, with 8B active parameters during prefill and 16B during decoding, as specified in the model card.
• A one-million-token context window, a controllable reasoning_effort setting from 1 to 100, and a multimodal vision encoder trained jointly with the language model from pre-training.
• An internal-test predecessor under deepseek-v4.1-flash-expires-on-0910, with a 20-concurrent-request cap that AiBattle_'s API testing notice captured and the early-test announcement described.
• Off-peak API prices of $0.003 per million cache-hit tokens, $0.15 for cache-miss plus write tokens, and $0.60 for output tokens. a DeepSeek pricing notification previewed the cut, while teortaxesTex's price post listed the rates and the two-times peak multiplier.
• A promised V4.1 Pro follow-on. eliebakouch noted the unusual confidence behind DeepSeek's claim that Flash already outperformed its updated Pro model.

First-party

• Terminal-Bench 2.1 Pass@1: 82.7% → 90.6%, +7.9 points, from V4 Flash to V4.1 Flash in DeepSeek's model card.
• DeepSWE v1.1 resolved: 54.4% → 74.2%, +19.8 points, from V4 Flash to V4.1 Flash in DeepSeek's model card.
• CyberGym Pass@1: 76.7% → 88.1%, +11.4 points, from V4 Flash to V4.1 Flash in DeepSeek's model card.
• AutomationBench Pass@1: 37.7% → 54.8%, +17.1 points, from V4 Flash to V4.1 Flash in DeepSeek's model card.
• HLE with tools Pass@1: 51.5% → 63.9%, +12.4 points, from V4 Flash to V4.1 Flash in DeepSeek's model card.

Third-party evaluators

• A frozen four-task comparison's visual-task success: 0/4 → 3/4, +75 points, from V4 Vision Exp to V4.1 Flash in a four-model test report screenshot.
• Aikido's vulnerability-agent tool calls per turn: 1.70 → 1.79, +0.09 calls per turn, in the Aikido research screenshot, which also described the newer model as more persistent when its initial scope came up empty.

Customer-reported

• Early reported output speed: roughly 1× → 3×, about +200%, from V4 Flash to V4.1 Flash in teortaxesTex's early speed test.

The same first-party table puts V4.1 Flash behind V4 Pro on several base-model measurements:

• SimpleQA-Verified: 55.2% → 42.3%, -12.9 points.
• LongBench-V2: 51.5% → 45.2%, -6.3 points.
• HLE without tools: 42.7% → 36.8%, -5.9 points.

Those comparisons come from DeepSeek's model card, which labels them base-model evaluations under the same internal settings.

Harness choice also moves the result. an engineering-evaluation screenshot reported a 16.77-point drop when the model was switched into DSH, plus functional correctness of 78 against Qwen 3.8 Flash's 82.

DeepSeek's technical report and model card answer the pre-release question of whether V4.1 was a thin post-training refresh: the company describes a from-scratch, 45T-token multimodal pre-training run, with sparse attention trained at 64K tokens and context extended to 1M at 34T tokens.

• CED cache projection: decoder global KV comes from the final encoder hidden states, rather than each decoder layer retaining its own global KV.
• SWA Bounded Replay: missing sliding-window states are rebuilt from the most recent window instead of being persisted to SSD, yielding about one eighth of V4 Flash's persistent KV footprint.
• CSA2 and FP4 cache: static Full, Reindex, and Reuse attention modes share index work across layers; FP4 KV caching brings global storage to 890 bytes per token, about four times smaller than V4 Flash and 437 times smaller than V1.
• MoE and memory: each layer has one shared expert and 384 routed experts, six selected per token, plus 196B-parameter Engram conditional memory accessed by token lookup.
• Vision: DeepSeek-ViT uses 2D-RoPE and 3×3 pixel-unshuffle downsampling, then a two-layer MLP projector combines visual and text embeddings from the start of pre-training.

The official chart uses maximum reasoning effort, temperature=1.0, and top_p=0.95. Code-agent results use Minimal DeepSeek Harness with a 1M-token context, while visual-agent tasks use Claude Code with 512K context; Agent's Last Exam and AutomationBench use their official scaffolds.

• In teortaxesTex's vision-iteration run, the agent rendered batches, built a contact sheet, and used the images to identify defects in its own procedural worlds.
• teortaxesTex's render test observed it modifying a render to test scene hypotheses and switching rendering modalities to make the scene more legible to itself.
• One long session reached 419 steps, 155M input tokens, and nearly 0.5M output tokens in about an hour; teortaxesTex's session report said the visual result still looked poor.
• The bridgemindai BridgeBench run took longer to finish than Fable 5.1 and GPT 6 Astra despite its high streaming rate, though its lava lamp was judged better than Muse Spark 1.3 and Gemini 3.8 Flash.

• Tencent's WorkBuddy and CodeBuddy, plus OpenCode, are listed as fully integrated official partners in the release announcement. OpenCode's availability post shows the model selector in OpenCode Go.
• The production DeepSeek API model name is deepseek-flash; deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1 Flash for compatibility, according to the official announcement.
• Requests addressed to deepseek-v4-pro will route to V4.1 Flash and receive Flash billing after 12:00 Beijing time on September 14, until V4.1 Pro arrives. aibuilderclub_'s rollout post recorded the interim replacement policy.
• DeepSeek published the weights under MIT and a local inference path that includes weight conversion instructions. Its release note says large-scale deployment discussions assume a 2,000-GPU cluster with storage, while AIBase's launch coverage reports the product merged the prior fast, expert, and image-recognition modes.
• Community launch threads collected the original Chinese announcement, the Hugging Face weights, and the technical report in one place, including this LocalLLaMA post.