DeepSeek releases V4.1-Flash fast reasoning model

On Thursday 2026-09-10, DeepSeek formally released DeepSeek-V4.1-Flash — "the smallest model in our new architecture family, with native visual understanding" — via an announcement on its news page, a release note on the API docs, MIT-licensed weights on Hugging Face (repo created 02:17 UTC), and a 1.81 MB technical report ("DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression"). Key facts:

1. Model specs: 552B backbone-parameter MoE with a Causal Encoder–Decoder (CED) architecture — 40 Transformer layers (20-layer causal encoder + 20-layer decoder) that activate only 8B parameters per token during prefill and 16B during decode. Total safetensors footprint 763B params (~510 GB, 48 shards), FP4+FP8 mixed precision, MIT license.
2. Efficiency story: global KV cache compressed to 890 bytes per token (~1/4 of V4-Flash), persistent SSD footprint ~1/8 of V4-Flash via SWA Bounded Replay, FP4 main KV caching (E2M1), Compressed Sparse Attention 2 (CSA2) with a Hierarchical Sparse Indexer, Single-Pass mHC, Engram conditional memory (196B params, sparse token-based lookup), and DSpark speculative decoding.
3. Multimodal: DeepSeek-ViT vision encoder (trained from scratch, 2D-RoPE, 3×3 pixel-unshuffle) — native image+text input, text output; image support that previously existed only in the experimental V4-Flash-Vision-Exp is now built in.
4. Training: from scratch on a 45T-token multimodal corpus; sparse attention trained at 64K sequence length with context extended to 1M tokens; post-training follows SFT → RL → on-policy distillation (OPD) with the changes in the data pipeline (large-scale automated synthesis of agent tasks/environments).
5. API + pricing: live immediately as deepseek-flash; peak/off-peak pricing — $0.30/$1.20 per 1M input/output tokens peak, $0.15/$0.60 off-peak, cache reads $0.006/$0.003 — effective 04:00 UTC Sept 10. V4-Flash and V4-Flash-Vision-Exp retired; their model IDs temporarily route to V4.1-Flash for compatibility.
6. V4-Pro retirement: from 04:00 UTC Sept 14, 2026, all deepseek-v4-pro requests route to V4.1-Flash at Flash rates (~70% output-price cut vs V4-Pro's $3.96/M), until V4.1-Pro launches.
7. Partner ecosystem: Tencent's WorkBuddy (incl. CodeBuddy) and the open-source coding agent OpenCode are named official partners with immediate V4.1-Flash support; Cambricon achieved "Day 0" silicon support (per Caixin); DeepSeek said it will work with the open-source community on inference support (vLLM/SGLang paths published on the model card).
8. Context: the same day, Anthropic's threat-intelligence report named DeepSeek among seven China-based labs it says ran distillation campaigns against Claude; Caixin and Reuters also noted DeepSeek has hired Citic Securities for a potential STAR Market IPO.

• Before: DeepSeek's V4 generation shipped in April 2026 (V4-Pro 1.6T/49B active; V4-Flash 284B/13B active, 1M context), with V4-Flash-Vision-Exp (Aug 21) as the only multimodal option. V4-Flash was the "fast" tier but still activated 13B params per token on a classic decoder-only shape, with a KV cache DeepSeek now characterizes as 4× larger per token than V4.1-Flash. The open-weight "flash-class" agentic bar was set by V4-Flash (Terminal-Bench 2.1: 82.7 per DeepSeek) and by Alibaba's Qwen 3.8 line.
• Change: DeepSeek replaced the fast tier with a new architecture family: CED asymmetric prefill/decode (8B/16B active), ~4× KV-cache compression, SSD footprint ~1/8, native vision, controllable reasoning effort (1–100), speculative decoding (DSpark) — all MIT-licensed, plus API price cuts (output $1.20/M peak vs $3.96/M for V4-Pro). It simultaneously retired V4-Flash/Vision-Exp and scheduled V4-Pro's effective retirement for Sept 14, making the "smallest model in the family" the de facto flagship for the interim.
• After: As of the window's end, deepseek-flash is the center of DeepSeek's API offering; V4-Pro is being rerouted to Flash (from Sept 14); the open-weight agentic frontier has a new cost/latency reference point; the community is already building quantizations (76), finetunes (14), GGUF ports (antirez/DwarfStar) and multi-GPU serving recipes; and a V4.1-Pro is pending to restore a two-tier lineup.

• Asymmetric Causal Encoder–Decoder (CED): 40 layers total — a 20-layer causal encoder followed by a 20-layer decoder. The decoder's global KV cache is projected from the final encoder hidden states instead of being derived per decoder layer, so the model activates only 8B params/token at prefill (input-heavy, agent-serviceable workloads) and 16B at decode.
• KV-cache compression stack (the headline engineering):
  • SWA Bounded Replay reconstructs missing sliding-window attention KV states by replaying only the most recent n_win tokens, eliminating SSD persistence of SWA KV → persistent footprint ~1/8 of V4-Flash.
  • CSA2 (Compressed Sparse Attention 2): each attention layer gets one of three static modes (Full/Reindex/Reuse) sharing main KV and indexer K across layers and reusing Top-K sparse-attention indices; a Hierarchical Sparse Indexer restricts later indexing layers to a candidate pool from the first Full layer, bounding cost independent of context length. Config confirms: kv_source_layer_ids [2,8,14,20], index_source_layer_ids [2,8,14,20,24,28,32,36], candidate_source_layer_id 20, candidate_topk_blocks 2048, index_topk 512.
  • FP4 main KV caching (E2M1 format, one E4M3 scale per 16 channels) → global KV footprint 890 bytes/token, ~1/4 of V4-Flash and ~437× smaller than DeepSeek-V1 per the model card figure.
• Compute extras: Single-Pass mHC (revised residual-stream mixing), Engram conditional memory (196B params across layers 1 and 14 — config: ~384M embedding rows per layer, 16M vocabulary, n-gram lookup up to 4), and DSpark speculative decoding (semi-autoregressive draft head, 3 next-token-prediction layers at 37–39, 128 draft experts, block size 5).
• Multimodal: DeepSeek-ViT encoder (2D-RoPE, 3×3 pixel-unshuffle downsampling) + two-layer MLP projector; visual and text embeddings processed jointly from the start of pre-training.
• Training: 45T-token multimodal corpus pretrained from scratch; sparse attention trained at 64K; context extended to 1M at 34T tokens (config: max_position_embeddings 1,048,576, yaRN factor 16 from a 64K base). Post-training: SFT → RL → OPD, unchanged recipe; the gains come from the data pipeline — large-scale automated synthesis of agent tasks/environments with progressive scaling of data, tasks and rollouts.
• Reasoning control: continuously controllable integer reasoning_effort (1–100) trading inference cost for accuracy; thinking mode default-on with a non-thinking option; B.AI documents max output 384K tokens and tool-use features (function calling, JSON output, context caching, image-bearing tool output).
• Serving/format: no Jinja chat template; a self-contained Python reference implementation (encoding.py) covers multi-turn, tool calls, thinking, numeric reasoning effort, mid-conversation system messages and interleaved images; deepseek-recipe (Rust + Python bindings) is the maintained prompt-format toolkit (GitHub, pushed release day).

• Open-weight frontier reasoning at a new cost/latency point: DeepSeek's claim — that the smallest member of the V4.1 family outranks its own 1.6T flagship on per-task cost, speed and benchmark totals — plus MIT licensing, makes it the strongest open-weights counter to closed-lab pricing this week, directly entering the framing of the week's pacing debate (S15, S22, S36) as evidence that "competition, not pause" continues in open weights.
• Architecture-led, not just benchmark-led, release: the differentiator is serving economics for long-running agents (expensive prefill, huge KV caches, cache-hit charges, 1M-token contexts) — KDnuggets and heise both frame the architecture, not the scores, as the story. This is a template for the next generation of "agent-grade" open models.
• Flag-ship substitution inside the window: DeepSeek effectively retired its flagship (V4-Pro) on Sept 14 and rerouted all traffic to the new model — an aggressive operational bet that the efficiency model is now the better product, with ~70% output-price cut for V4-Pro users.
• The open agent-stack race: official partner integrations with WorkBuddy/CodeBuddy (Tencent) and OpenCode, Cambricon Day-0 silicon support, and rapid community quantization (76 repos) show the open-weight agent ecosystem pivoting to this model within days — the same week Salesforce, Microsoft and OpenAI were pushing proprietary agent platforms (S26, S30, S29).

• Low-cost, long-context agent backends: an MIT-licensed 1M-context multimodal reasoning model whose per-token active compute (8B/16B) and 890 B/token KV cache make sustained agent sessions (which burn cache reads) materially cheaper to operate — cache-hit pricing ($0.003–0.006/M) directly targets agent cost structure.
• "Flash-class" self-hosting on commodity-to-mid clusters: 48-shard FP4/FP8 weights enable FP8 inference on H100-class nodes; community work shows viable 4× RTX PRO 6000 (96 GB) or SSD-streamed single-node serving paths; Q2 GGUF builds run on 128 GB Macs with SSD streaming (DwarfStar).
• Controllable reasoning in production: a continuous effort dial (1–100) lets builders tune latency vs. accuracy per request class — a practical latency-lowering mechanism for interactive/agent apps that closed APIs mostly expose as coarse tiers.
• Open multimodal + agentic convergence: native vision in an open model with strong chart/UI/terminal agent scores (Chartography 78.9, Terminal-Bench 2.1 90.6 per DeepSeek) lets teams build self-hosted "see-and-do" agents without closed-vendor lock-in.

1. When does V4.1-Pro ship, and at what price/positioning relative to the retired V4-Pro? (DeepSeek gave no date.)
2. Do independent re-runs reproduce the agentic benchmark leads (Terminal-Bench 2.1 90.6, DeepSWE 74.2, CyberGym 88.1) under standard harnesses, and does the verbosity tax persist across real agent workloads?
3. How large is Engram's real contribution — ablation-style evidence is absent so far; does the 196B memory improve long-context retrieval or mostly inflate download size?
4. What are actual Day-0 Cambricon/domestic-HW performance numbers, and how real is the domestic deployment path under current export controls?
5. Will the reasoning-effort dial (1–100) become an open standard across providers (B.AI, Novita, Fireworks, Baseten, Databricks, DigitalOcean, Inco, LithosAI, Makora already host it)?
6. Does the V4-Pro reroute hold — will customer churn or quality complaints force DeepSeek to keep a Pro tier alive before V4.1-Pro ships?

• Sept 14 (in-window): V4-Pro rerouting takes effect 04:00 UTC — every deepseek-v4-pro call now runs V4.1-Flash at Flash rates; watch for quality complaints or migration breakage.
• Weeks ahead: V4.1-Pro announcement (no date given); first independent full benchmark re-runs (Vals/AA updates, possible Terminal-Bench/DeepSWE re-publication); maturation of vLLM/SGLang/llama.cpp support and 76+ quantizations; possible ModelScope/domestic-Cambricon deployments at scale.
• Quarter ahead: DeepSeek STAR Market IPO prep continues (Citic Securities); the V4.1 architecture likely scales up (Pro) and possibly down (community "Lite" requests) — DeepSeek has not committed to a small local model despite community pressure on HF discussions.

DeepSeek-V4.1-Flash is a release about serving economics wearing a benchmark story. The headline — an open MIT-licensed model that makes its own flagship redundant — matters less than the architecture underneath it: asymmetric prefill/decode, a 4× smaller KV cache, speculative decoding and a continuous reasoning-effort dial, all aimed at the real cost center of the agent era (long contexts, cache reads, per-task token burn). Independent testers confirmed the agentic gains and flagged the verbosity tax and the hard-reasoning gap. In a week dominated by agent-safety incidents and pleas to slow down, DeepSeek's answer was to make frontier-class agent capability cheaper and fully open — a forceful, market-level counterpoint that the pacing debate will have to contend with.

Evidence-status summary: CONFIRMED (release, date, specs — primary sources + repo inspection); COMPANY CLAIM (benchmark superiority, "beats V4-Pro," flag-ship-outperformed framing); INDEPENDENT EVIDENCE (Vals AI, Artificial Analysis, heise/BenchLM aggregation, provider benchmarks, token-verbosity caveats); INTERPRETATION/PREDICTION as labeled inline.