# DeepSeek V4.1 Flash DeepSeek V4.1 Flash is [[Deepseek]]'s [[AI Open Weight Models|open-weight]] (MIT) multimodal [[AI Mixture of Experts (MoE)|MoE]] model, released 10 September 2026, one week after [[GPT-6 Astra]] and [[Claude Fable 5.1]]. The ".1" undersells it. This is a new architecture, twice the parameters of [[DeepSeek v4|V4 Flash]], native vision, and benchmark numbers that beat the much larger V4 Pro across the board. DeepSeek retired V4 Pro four days after launch and now routes Pro requests to this model at Flash prices. The tech report title says where the effort went: *Pushing the Limits of KV Cache Compression*. ## What it is | | V4 Flash | V4 Pro | **V4.1 Flash** | |---|---|---|---| | Backbone parameters | 284B | 1.6T | **552B** (+196B Engram memory) | | Active per token | 13B | 49B | **8B prefill / 16B decode** | | Context | 1M | 1M | **1M** (max output 384K) | | Vision | experimental variant | no | **native** | | Global KV cache per token | ~3.5 KB | | **890 bytes** | Trained from scratch on 45T multimodal tokens. Reasoning effort is a **continuous integer from 1 to 100** rather than a handful of named levels (the API maps `low`/`high`/`max` onto it; there is no real medium, which people complain about). Speaks the OpenAI Chat Completions, Responses, and [[Anthropic]] API formats natively. A DeepSeek Harness exists and the model was evaluated inside [[Claude Code]], Codex, [[OpenCode]], and Pi as well. ## The architecture, briefly Every piece of this is aimed at one thing: ingest a lot of input tokens cheaply and keep the cache tiny. That is the profile of agentic work, where the model re-reads a large context on every turn and writes comparatively little. - **Causal Encoder-Decoder (CED).** A 40-layer transformer split into a 20-layer causal encoder and a 20-layer decoder. The decoder's global [[AI KV Cache|KV cache]] is projected from the encoder's final hidden states rather than computed per decoder layer. That's how prefill gets away with 8B active parameters - **Compressed Sparse Attention 2 (CSA2).** Each attention layer runs in one of three static modes (Full, Reindex, Reuse) so layers share main KV and indexer keys and reuse top-k sparse indices. A hierarchical indexer in the decoder bounds indexing cost independently of context length. Successor to the DSA introduced in V3.2 and used in V4 ([[AI Sparse Attention]]) - **FP4 main KV cache** (E2M1 with one E4M3 scale per 16 channels). With the above: 890 bytes per token, about a quarter of V4 Flash and 437× less than DeepSeek V1. A full 1M-token context is roughly 900 MB of cache - **SWA Bounded Replay.** Sliding-window KV isn't persisted; it's rebuilt by replaying the last window of tokens. Persistent cache footprint drops to about an eighth of V4 Flash, which is what makes context checkpoints on SSD cheap - **Engram conditional memory.** 196B parameters of n-gram memory accessed by token lookup, separate from the backbone. Works like a database read per token, so it can live on fast SSD instead of RAM. Same idea as Qwen 3.8 Flash Next's engram - **DSpark speculative decoding.** Semi-autoregressive drafts with confidence-scheduled verification ([[AI Speculative Decoding]]) - **Single-Pass mHC** residual mixing, 1 shared + 384 routed experts with 6 routed active per token, and a from-scratch vision encoder (DeepSeek-ViT, 2D-RoPE, 3×3 pixel unshuffle) One [[Hacker News]] reading I agree with: V4.1 Flash is optimised for agentic tool calling at the expense of world knowledge and long-context recall. The base-model table backs that up: MMLU-Pro and code improved over V4 Pro, SimpleQA and LongBench-V2 did not. More total parameters, fewer active ones, tiny cache: they wanted the capability without paying for it in FLOPs per token. Post-training is the standard SFT → RL → on-policy distillation recipe "without algorithmic modifications"; the changes are all in the data pipeline (large-scale synthesis of agent tasks and environments). That sentence is worth keeping next to the [[DeepSeek v4|V4-Flash-0731]] lesson: the gains keep coming from post-training data, not from tricks in the loss function. ## Benchmarks DeepSeek's own table at max effort, against the models people actually compare it to: | Benchmark | [[Claude Opus 5|Opus 5]] | [[GPT-5.6|GPT-5.6 Sol]] | [[Kimi K3|K3]] | GLM-5.3 | V4 Pro | **V4.1 Flash** | |---|---|---|---|---|---|---| | Terminal-Bench 2.1 | 89.1 | 88.8 | 88.3 | 88.2 | 87.9 | **90.6** | | Terminal-Bench 4.0 | **51.8** | 39.9 | 12.6 | 37.9 | 12.4 | 31.2 | | DeepSWE v1.1 | 74.0 | 73.0 | 67.5 | 66.9 | 62.7 | **74.2** | | NL2Repo-Bench | **75.3** | 56.8 | 58.0 | 58.0 | 61.5 | 64.0 | | CyberGym | | 84.5 | 80.0 | 84.5 | 83.3 | **88.1** | | SEC-Bench Pro | | **74.3** | | | 56.4 | 62.8 | | AutomationBench | 50.3 | 45.8 | 46.7 | 48.8 | 43.2 | **54.8** | | Agent's Last Exam | 28.6 | 26.7 | 27.6 | 28.5 | 25.7 | **31.8** | | GPQA Diamond | 93.4 | **94.1** | 92.9 | 88.1 | 92.4 | 90.9 | | HLE (no tools) | **56.3** | 44.5 | 43.5 | 42.0 | 42.7 | 36.8 | | Codeforces rating | | | | | 3348 | **3471** | Read the table, not the headline. V4.1 Flash leads on the agentic benchmarks that reward many cheap tool calls (Terminal-Bench 2.1, DeepSWE, AutomationBench, Agent's Last Exam, CyberGym) and trails badly on the ones that reward deep reasoning per step (Terminal-Bench 3.0 and 4.0, HLE, ProgramBench, SEC-Bench Pro). Against Opus 5 on Terminal-Bench 4.0 it's 31.2 versus 51.8. That's the same model, and both numbers are true. The scaffold table is also worth a look: DeepSWE resolved rate ranges from 65.5% (OpenCode) to 74.2% (mini-SWE) for the same weights. Nine points from the harness alone. ## Pricing Per million tokens, new prices effective 10 September 2026: | | Off-peak | Peak | |---|---|---| | Input, cache hit | **$0.003** | $0.006 | | Input, cache miss | $0.15 | $0.30 | | Output | $0.60 | $1.20 | Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Everything else is half price. Two things people got wrong in the thread. "They cut prices" and "they quadrupled prices" are both circulating; the truth is that cache-hit input got cheaper (halved) while cache-miss input and output roughly doubled to quadrupled versus V4 Flash's $0.14 / $0.28. Whether you pay more depends entirely on your cache hit rate, and on agentic loops that rate is above 99%. One commenter priced a real 447-turn coding task (1.0M input, 165K output, 36.5M cache reads) on both models. [[GPT-6 Astra]] at $10 / $50 / $1: about **$55**, two thirds of it cache reads. V4.1 Flash off-peak: about **$0.50**, of which cache reads are a rounding error. A 100× gap on the workload shape that matters. The follow-up question is a good one: at $0.003 per million cached tokens, the cost of moving the context over the network is starting to matter more than the cost of the model reading it. ## Reception The launch thread hit 1,012 points, the leak thread 417. What people actually said: - **Speed is the story.** 300 to 500 tokens/s in beta, about 150 on [[OpenRouter]] after launch. Still roughly double V4 Flash. "Gemini Flash is no longer the fastest option" - **It runs at home, barely.** 552B in ~FP4 is about 306 GB, plus 204 GB of FP8 Engram that can sit on SSD. Practically that's three DGX Sparks or four RTX PRO 6000s, and it no longer fits the 2 × 128 GB unified-memory setups that ran V4 Flash. People who bought hardware for the old Flash are annoyed. "Not really flash anymore" got the right reply: Flash means fast, and this one is faster - **Fewer refusals.** The recurring reason people reach for Chinese models: it will review your own codebase for an out-of-bounds bug, write a UDP packet handler for a game, or download a ROM you own, where Astra and Claude refuse. One person routes requests to Flash automatically whenever Astra trips a cybersecurity flag - **It's a workhorse, not a planner.** The pattern that keeps showing up, here and in the [[GPT-6 Astra]] threads: an expensive model plans and reviews (Opus, Fable, Astra), Flash implements. "DeepSeek IS the workhorse" for people building software factories where Anthropic pricing is a non-starter - **Reasoning is too eager.** "Low" is close to off, "high" is close to "max", and max thinks forever with a lot of "but wait". Several people moved to GLM-5.3 Flash for the more usable effort curve. Simon Willison's pelican run confirms OpenRouter exposes seven levels that map onto three - **Language drift.** The consumer app answers in Chinese to English prompts often enough that people uninstalled it. The API doesn't do this. Third-party hosts with lower quants also produce the infinite loops and invalid tool calls people blame on the model - **The tech report.** Universal praise for a report full of architectural detail, with the inevitable comparison to a Fable system card that's mostly safety and model welfare. And the inevitable Anthropic distillation accusation in reply, which the thread didn't resolve and I won't either ### Retiring Pro by rerouting it DeepSeek's announcement: "all requests to the Pro model will be routed to V4.1 Flash and billed at Flash's price." Postponed once (to 14 September) and then done. Users loved the price and hated the mechanism, and the second group is right. If you validated a workflow on V4 Pro, your production traffic now runs on a different model with a different architecture and you didn't opt in. Simon Willison's version: API providers should lean toward *not* swapping models on paying customers, however much better the new one is. The open weights soften it (Pro stays available on [[OpenRouter]], Together, and anyone else hosting it), but it's a good reason to pin models by exact name and to keep evals that would catch a silent swap. ### The uncensored variant Within a day, dealign.ai published `DeepSeek-V4.1-Flash-UNCENSORED-FP8`, an abliterated checkpoint with the refusal circuitry removed at the weight level. Their numbers: HarmBench compliance from 1.6% to 100% at max effort, MMLU down 4.2 points overall and 1.1 points outside the ethics subjects. Interesting detail in their table: the *base* model refuses far more at max effort than with reasoning off (42.8% compliance drops to 1.6%), because reasoning surfaces the safety concern before the answer. The HN thread on it asked whether abliteration means alignment is a lost cause. For open weights, refusal training has always been a speed bump, not a wall. What's new is the turnaround: hours, on a 552B model. ## What I take from it - **The "flash beats last quarter's pro" cadence is now the norm** for Chinese labs (GLM 5.3 Flash over 5.2, now this). Plan your model choices around it: don't build a moat on a Pro model's specific behaviour - **Architecture is the moat DeepSeek is building**, and it's aimed squarely at agent economics: cheap prefill, tiny cache, fast decode. The frontier labs compete on capability per task; DeepSeek competes on cost per task, and on the workloads in the scaffold table it now wins both - **Pin model names and keep evals.** The Pro reroute is a preview of how every provider will behave when the cheaper model wins their internal tests - **Use it where it's strong.** Implementation, triage, bulk agentic loops, vision-based QA, security review that US models refuse. Keep the deep single-step reasoning (Terminal-Bench 4.0 shape) on Opus or Fable - **Watch what you send.** DeepSeek's own API trains on prompts. Use a zero-data-retention host for anything that matters ## References - Model card and tech report: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash - API announcement: https://api-docs.deepseek.com/news/news260910 - Pricing: https://api-docs.deepseek.com/quick_start/pricing - Thinking mode and effort mapping: https://api-docs.deepseek.com/guides/thinking_mode/ - DeepSeek Harness: https://deepseek.com/harness/en/ - deepseek-recipe (prompt encoding toolkit): https://github.com/deepseek-ai/deepseek-recipe - Announcement thread (X): https://x.com/deepseek_ai/status/2097930608790167907 - [[Hacker News]]: launch (1,012 points) https://news.ycombinator.com/item?id=49639090 · pre-launch leak and Pro reroute (417 points) https://news.ycombinator.com/item?id=49624603 · uncensored variant https://news.ycombinator.com/item?id=49654387 - Uncensored checkpoint: https://huggingface.co/dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 - OpenRouter listing: https://openrouter.ai/deepseek/deepseek-v4.1-flash ## Related - [[Qwen3.8-Flash-Next]] - [[Deepseek]] - [[DeepSeek v4]] - [[DeepSeek V3]] - [[AI Mixture of Experts (MoE)]] - [[AI Open Weight Models]] - [[AI KV Cache]] - [[AI Sparse Attention]] - [[Sparse AI Models]] - [[AI Speculative Decoding]] - [[AI Reasoning Models]] - [[AI Inference]] - [[Context Window]] - [[Knowledge Distillation]] - [[AI Alignment]] - [[Kimi K3]] - [[GLM-5.2]] - [[GPT-5.6]] - [[GPT-6 Astra]] - [[Claude Opus 5]] - [[Claude Fable 5.1]] - [[Claude Code]] - [[OpenCode]] - [[OpenRouter]] - [[HuggingFace]] - [[llama.cpp]] - [[Simon Willison]] - [[Hacker News]] - [[Large Language Models (LLMs)]]