DeepSeek V4.1: Release Date, What Changed From V4, and Benchmarks

What Is DeepSeek V4.1?

Last updated November 9, 2026.

DeepSeek V4.1 is DeepSeek's newest open-weight model generation, and as of November 9, 2026 it contains one model: DeepSeek V4.1 Flash. DeepSeek released it on September 10, 2026 and shipped an update, V4.1 Flash 1017, on October 17. There is no V4.1 Pro. DeepSeek's API still serves V4-Pro-0813 under the deepseek-v4-pro name. V4.1 Flash is a new design, not a retrain of V4. It is a mixture-of-experts model with a 552B backbone plus 196B of Engram conditional memory, and it activates 8B parameters per token during prefill and 16B during decode. It reads images natively, handles a 1M-token context, and ships under the MIT license. In DeepSeek's own tables it beats V4-Pro-0813 on agent benchmarks, for example 74.2 vs 62.7 on DeepSWE v1.1, while V4-Pro keeps the lead on knowledge tests such as GPQA Diamond. DeepSeek's API bills it as deepseek-flash at $0.15 per million input tokens and $0.60 per million output tokens off-peak.

DeepSeek V4.1 Release Date and Timeline

DeepSeek released V4.1 Flash on September 10, 2026. The Hugging Face repository went up at 02:17 UTC that day, and the new API prices took effect at 04:00 UTC. The same announcement retired V4 Flash and announced a plan to retire V4 Pro, which DeepSeek later reversed. Five weeks later, on October 17, the 1017 update replaced the September checkpoint everywhere DeepSeek serves it.

The "V4.1" name covers one model so far. When people search for DeepSeek V4.1, the thing they can call is V4.1 Flash, and on DeepSeek's own API that now means the 1017 checkpoint. OpenRouter lists the two checkpoints separately as deepseek/deepseek-v4.1-flash and deepseek/deepseek-v4.1-flash-1017. For the model as a product (pricing across providers, Claude Code and Codex setup, cache economics), see the DeepSeek V4.1 Flash page. For the October update on Morph, see DeepSeek V4.1 Flash 1017.

Is There a DeepSeek V4.1 Pro?

No. DeepSeek has not released a V4.1 Pro as of November 9, 2026, and it has published no V4.1 Pro specs, prices or benchmarks. Four primary sources agree:

• The launch post expected one. It said all deepseek-v4-pro requests would route to V4.1 Flash at V4.1 Flash rates from September 14, and that "this will continue until V4.1-Pro launches."
• The reroute was cancelled. DeepSeek's change log says that "in response to user demand," it will continue providing API services for V4 Pro after September 14, 2026, with billing unchanged.
• The Pro API name still serves V4. DeepSeek's pricing page lists the model behind deepseek-v4-pro as DeepSeek-V4-Pro-0813.
• No weights exist. The newest V4-family repositories in the deepseek-ai Hugging Face organization are V4.1-Flash-1017 and V4.1-Flash. There is no V4.1-Pro repository.

DeepSeek does describe V4.1 Flash as "the smallest model in our new architecture family" and says the architecture is designed for "scaling to larger models." A larger sibling is implied. It has no date. If you need a Pro-sized DeepSeek model today, the choice is between the two models DeepSeek actually serves:

V4-Pro-0813 still wins on knowledge. In DeepSeek's V4.1 launch table it scores 92.4 on GPQA Diamond to V4.1 Flash's 90.9, and 42.7 to 39.1 on the text-only subset of Humanity's Last Exam. At the base-model level, V4-Pro-Base leads SimpleQA-Verified 55.2 to 42.3 and LongBench-V2 51.5 to 45.2. On the agent and coding rows of DeepSeek's launch table, V4.1 Flash is ahead, at 3.3x lower output cost off-peak. Send agent loops and coding to V4.1 Flash. Test V4-Pro on factual recall and long-document reading before you assume Flash is enough.

DeepSeek V4.1 vs V4: What Changed

The jump from V4 to V4.1 is an architecture change. DeepSeek's V4-Flash-0731 update "keeps the same model architecture and size" as the V4 Flash preview and "was only re-post-trained." V4.1 Flash was trained from scratch on a new design that DeepSeek says is built for "a higher capability ceiling, faster inference, higher throughput, and scaling to larger models." For the original V4 Pro and V4 Flash, including serving benchmarks and the V4-era pricing history, see the DeepSeek V4 guide.

The changes that matter in practice:

1. Prefill is cheap by design. The model activates 8B parameters per input token and 16B per output token. The model card says this substantially improves "cost efficiency for input-heavy agentic workloads," which is what a coding agent is: it rereads a large context every turn and writes a short reply.
2. The KV cache shrank. DeepSeek says V4.1 Flash needs 1/4 the HBM and 1/8 the SSD storage for its KV cache compared with the previous generation. Its reason is billing: "Cache-hit charges often account for a large share of agent costs."
3. Vision is native. V4 Flash was text only, and images went through the experimental V4-Flash-Vision-Exp. V4.1 Flash was pre-trained on image and text together from the start.
4. Prices went down. DeepSeek's change log says "API prices have been reduced accordingly" with the release. The current rates are in the API section.
5. The Pro tier did not move. V4.1 has no Pro model, so the Flash-vs-Pro choice is now a cross-generation choice: new-architecture Flash against old-architecture Pro.

DeepSeek V4.1 Architecture: CED, CSA2, Engram, and the FP4 KV Cache

Everything in this section comes from the DeepSeek-V4.1-Flash model card on Hugging Face, which DeepSeek titled "Pushing the Limits of KV Cache Compression." DeepSeek also links a technical report PDF from the card.

Causal Encoder-Decoder (CED)

V4.1 Flash is a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder. The decoder's global KV cache is projected from the encoder's final hidden states instead of being computed from each decoder layer's own hidden states. That split is where the two active-parameter numbers come from: 8B per token during prefill, 16B during decode. Each MoE layer has 1 shared expert and 384 routed experts, and 6 routed experts fire per token.

Compressed Sparse Attention 2 (CSA2)

CSA2 gives every attention layer one of three static modes: Full, Reindex or Reuse. Layers share the main KV and the indexer keys, and reuse Top-K sparse-attention indices instead of recomputing them. In the decoder, a Hierarchical Sparse Indexer limits later indexing layers to a candidate pool built by the first Full-mode layer, so the cost of the deeper indexers does not grow with context length.

FP4 KV cache and SWA Bounded Replay

The main KV cache is stored in FP4 (E2M1, with one E4M3 scale per 16 channels). Together with CSA2 that brings the global KV cache to 890 bytes per token, roughly a quarter of V4-Flash and, per the card, about 437 times smaller than DeepSeek-V1. At 890 bytes per token, a full 1M-token context holds under 1 GB of global KV. SWA Bounded Replay handles the sliding-window layers: instead of persisting their KV to SSD, it rebuilds missing sliding-window state by replaying only the most recent window of tokens. That cuts the persistent KV footprint to about 1/8 of V4-Flash.

Engram conditional memory

Engram is a separate 196B-parameter memory that the model reads through a sparse, token-based lookup. DeepSeek counts it outside the 552B backbone. For self-hosters this matters because the tables do not have to sit in GPU memory: vLLM's recipe exposes an Engram CPU-offload option (covered in the self-host section).

Other components

• Single-Pass mHC: revised residual-stream mixing with a fused Mega-mHC kernel.
• DSpark speculative decoding: semi-autoregressive draft generation with confidence-scheduled verification. vLLM's recipe notes that V4.1 dropped the MTP module, so DSpark is the only speculative method for this checkpoint.
• Vision: a DeepSeek-ViT encoder trained from scratch (2D-RoPE, 3x3 pixel-unshuffle downsampling) and a two-layer MLP projector turn images into embeddings that the model processes alongside text.

Training

DeepSeek pre-trained V4.1 Flash from scratch on 45T multimodal tokens. Sparse attention trained at a 64K sequence length, and the context was extended to 1M tokens at the 34T token mark. Post-training follows SFT, then RL, then on-policy distillation "without algorithmic modifications." DeepSeek says the substantive changes were in the data pipeline, through "large-scale automated synthesis of agent tasks and environments." The 1017 update adds reinforcement learning with self-feedback aimed at end-to-end project development.

DeepSeek V4.1 Benchmarks vs V4

Every number below is DeepSeek-reported. DeepSeek ran all instruct results at maximum reasoning effort (reasoning_effort=100) with temperature 1.0 and top_p 0.95. Code-agent benchmarks ran in the Minimal mode of DeepSeek Harness with a 1M-token context, except DeepSWE v1.1 (mini-SWE harness) and SEC-Bench Pro (Claude Code). No independent lab has published a reproduction of these rows.

V4.1 Flash leads 14 of the 16 rows, one of them (MathArena Apex) by 0.3 points. V4-Pro-0813 keeps GPQA Diamond and Humanity's Last Exam without tools. The biggest gains are on the newer, harder terminal suites: on Terminal-Bench 3.0, V4.1 Flash scores 30.0 against V4-Pro's 11.8, and on Terminal-Bench 4.0, 31.2 against 12.4. On the older Terminal-Bench 2.1, every model in the table is above 82, so the gap there is 2.7 points.

Base models: where V4.1 gains and where it does not

DeepSeek also published base-model scores from its internal framework under identical settings. It treats scores within 0.3 of each other as equivalent.

The base table explains the instruct table. V4.1-Flash-Base beats V4-Flash-Base on factual recall (42.3 vs 30.1 on SimpleQA-Verified) and leads all three on code, but V4-Pro-Base still knows more and reads long documents better. V4.1 also scores lowest of the three on MGSM, the multilingual grade-school math set (80.2). The base-model gains on code are a few points, so most of the agentic jump in the instruct table likely comes from post-training, where DeepSeek says its changes were "large-scale automated synthesis of agent tasks and environments."

The independent score: Artificial Analysis Intelligence Index v5

Artificial Analysis published Intelligence Index v5 on October 27, 2026. It scores DeepSeek V4.1 Flash 1017 (max) at 32, level with GLM-5.3 (max) at 32 and below Claude Haiku 5.5 (max) at 36. Kimi K3.1 Pro scores 46 on the same index. Artificial Analysis puts the two 32s at "26 points behind the leader." This is an aggregate of Artificial Analysis's own evaluations, not a re-run of DeepSeek's rows, and it covers 1017 only. Compare it only with other v5 scores, because earlier index versions use a different scale.

The 1017 Update: What Changed in DeepSeek V4.1 Flash 1017

DeepSeek-V4.1-Flash-1017 arrived on October 17, 2026. The model card calls it "a minor update" with "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks." The release post adds "faster task completion, and shorter outputs than the previous version," and says DeepSeek used reinforcement learning with self-feedback to improve end-to-end project work, from game creation to 3D modeling. It is tuned for specific harnesses: DeepSeek names DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and asks users to upgrade to DeepSeek Harness v0.2.1-rc.1 or later.

Every row goes up, but DeepSeek chose the rows. The 1017 table drops several benchmarks from the launch table, including GPQA Diamond, Codeforces, Terminal-Bench 2.1 and 3.0, CyberGym and SEC-Bench Pro, so there is no first-party 1017 number for those. DeepSeek's 1017 post also plots DeepSWE v1.1 pass rate against cost per task (77.0% at max effort, mini-swe-v2 harness), but the chart has no labeled dollar values, so we do not quote a cost.

The switch was automatic on DeepSeek's side. The API name deepseek-flash now serves 1017 at the same price, and the web and app models moved with it. On Morph, morph-dsv41flash and morph-dsv41flash-1017 route to the same deployment at the same price, as the DeepSeek V4.1 Flash 1017 page explains.

Harness Choice Moves DeepSeek V4.1 Scores

DeepSeek's model card runs the same model through eight agent scaffolds on the same two benchmarks. All runs use max effort, temperature 1.0, top_p 0.95, a 1M-token context limit and up to 500 steps per agent, with 8 samples per task on DeepSWE v1.1 and 3 on Terminal-Bench 2.1 (no network access).

Both headline numbers are the best scaffold in DeepSeek's own sweep. The 74.2 on DeepSWE v1.1 is the mini-SWE run; the 90.6 on Terminal-Bench 2.1 is DeepSeek Harness Minimal. In Claude Code the same model scores 69.8 and 88.0. In Codex it scores 65.6 and 84.1. The spread across scaffolds is 8.7 points on DeepSWE and 6.5 on Terminal-Bench 2.1, larger than several of the version-to-version gains in the tables above.

The 1017 Terminal-Bench 4.0 jump crosses harnesses

At launch, DeepSeek ran Terminal-Bench 4.0 in DeepSeek Harness Minimal mode. For 1017, its change log says "Terminal-Bench 4.0 was tested with the Claude Code framework." The 1017 table reprints 31.2 for the September model, the same figure as at launch, and DeepSeek does not say whether it re-ran that model in Claude Code. So the 11.1-point gain compares two harnesses, and part of it may be the harness. DeepSWE moved the same way, from mini-SWE at launch to Mini-SWE-V2 for 1017.

Outside runs add more noise. Moonshot's Kimi K3.1 Pro model card includes 1017 in its comparison tables and notes that its internal evaluation of DeepSeek-V4.1-Flash-1017 on Terminal-Bench-Science 0.1 yields 9.3, below the 15.7 it lists for the original V4.1 Flash. That is one competitor's run on one benchmark, but it shows the 1017 gains are not uniform outside DeepSeek's chosen suite.

The practical rule: measure V4.1 in the harness you will ship. DeepSeek's model card includes an evaluation folder with step-by-step instructions to reproduce DeepSWE v1.1 with both its dsh-minimal agent and the official mini-swe-agent, which is a good place to start a comparison against your own scaffold.

DeepSeek V4.1 Thinking Mode Rules

On DeepSeek's API, V4.1 Flash runs in thinking mode by default, at high effort. Non-thinking mode is one parameter away. The controls depend on which API format you call:

• Other effort names are remapped. DeepSeek maps minimal to low, medium and xhigh to high, and ultra to max.
• Sampling knobs are mostly ignored. Thinking mode does not support temperature, presence_penalty or frequency_penalty. On deepseek-flash, top_p only applies in thinking mode, with an effective range of 0.95 to 1.0; lower values are treated as 0.95. In non-thinking mode it is fixed at 1.0.
• Benchmarks used max. DeepSeek's published scores use maximum effort. If you leave the default at high, expect lower numbers than the tables above.
• FIM completion needs thinking off. DeepSeek's fill-in-the-middle completion (beta) works in non-thinking mode only.

Open weights: a 1 to 100 effort dial

The released weights expose a finer control than the API. The model card describes a continuously controllable reasoning effort, an integer from 1 to 100 that trades inference cost for accuracy, and DeepSeek's benchmark runs use 100. vLLM's recipe says that with no effort and no thinking key set, thinking is on at effort 50. One difference catches people moving from the API to a self-hosted server: vLLM's recipe says minimal and medium are rejected, while DeepSeek's API silently maps them. A client that sends medium works against api.deepseek.com and errors against your own vLLM.

Run the DeepSeek V4.1 Open Weights Yourself

Both checkpoints are on Hugging Face under the standard MIT License, covering the code and the weights: deepseek-ai/DeepSeek-V4.1-Flash and deepseek-ai/DeepSeek-V4.1-Flash-1017. The 1017 weights are also on ModelScope. The license's only condition is keeping the copyright and permission notice; it has no use-based restrictions.

No Jinja chat template

The repository ships no Jinja-format chat template, so a stock tokenizer template will not format prompts correctly. DeepSeek gives two replacements. The encoding folder holds a self-contained Python reference, encoding.py, with test cases for multi-turn chat, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages and interleaved images. For production, DeepSeek released deepseek-recipe, Rust libraries with Python bindings that convert Messages, Chat Completions and Responses API requests into V4.1 prompts and parse model output back, including streaming, thinking, tool calls and images. Inference, tool execution and HTTP are left to you.

vLLM

The vLLM recipe requires vLLM 0.31.0 or later and defaults to the 1017 checkpoint; it labels the September checkpoint "FP8 (0909)" for anyone reproducing earlier results. It describes the released checkpoint as MXFP4 routed experts with everything else MXFP8 block-quantized, about 511 GB on disk. By the recipe's numbers that fits one GB200 NVL4 tray (768 GB) at TP4, or one 8-GPU H200 node (1,128 GB) with room for KV cache. Its hardware selector lists H100, H200, B200, B300, GB200, GB300 and AMD MI300X through MI355X. The V4.1-specific flags in its GB200 command (an excerpt, not a complete command):

--tokenizer-mode deepseek_v41
--tool-call-parser deepseek_v41
--enable-auto-tool-choice
--reasoning-parser deepseek_v41
--engram-config '{"cpu_offload":true,"use_thp":true}'
--speculative-config '{"method":"dspark", ...}'

The Engram offload flag keeps the 196B memory tables in host RAM. The recipe adds that the DSpark drafter brings roughly 14B extra parameters to load. It also notes that its verified hardware results were run on the 0909 checkpoint and that the 1017 weights have not been separately re-run, and parts of the page disagree with each other (for example TP2 vs TP4 on Blackwell). Treat it as a starting point and check the full command on the page before you deploy.

SGLang

The SGLang cookbook states that V4.1 Flash support "has not shipped in an SGLang release yet, so a stock pip install sglang cannot serve it." NVIDIA GPUs (H200, B200, B300, GB300) use the preview image lmsysorg/sglang:dev-dsv41, and MI355X has its own ROCm image. Three details from the cookbook:

• Tool calls need the V4.1 detector. --tool-call-parser auto resolves to deepseekv41. V4.1 uses spaced DSML tool tags that the stock DeepSeek-V4 detector does not parse. If raw DSML markup shows up in content, the parser flag is missing.
• Two recipes. Low-Latency turns on DSpark with --speculative-algorithm DSPARK --speculative-dspark-block-size 5. High-Throughput turns speculation off, because at large batch sizes the fixed step cost stops paying for itself.
• Output is not bitwise stable. The cookbook says output is not bitwise stable across batch composition today, and --enable-deterministic-inference is refused on this backend. Plan evals around that.

NVIDIA Dynamo

NVIDIA's Dynamo recipe deploys V4.1 Flash on eight B200, GB200 or H200 GPUs, either aggregated as two TP4 workers or disaggregated as one TP4 prefill worker and one TP4 decode worker. Its SGLang targets are GB200 only; H200 and B200 run on vLLM. It says aggregated serving enables DSpark while disaggregated SGLang serving runs without speculative decoding; the SGLang cookbook says its DSpark recipes keep speculation in prefill/decode mode. The two docs disagree, so measure both. The Dynamo page also warns that with disaggregated SGLang, an HTTP 200 can still carry content: null and zero completion tokens when the KV transfer fails, so alert on empty completions as well as status codes.

Recommended settings

The 256K figure is the model card's recommended max_tokens setting for your own deployment. It is a different number from DeepSeek's API cap of 384K output tokens. If you self-host and cap output much lower, long thinking runs will truncate.

DeepSeek V4.1 API: First-Party Pricing and Morph

DeepSeek serves V4.1 Flash as deepseek-flash at https://api.deepseek.com (OpenAI format) and https://api.deepseek.com/anthropic (Anthropic format). The API supports tool calls, JSON output, the OpenAI Responses API, chat prefix completion and FIM completion (both beta), with a limit of 2,500 concurrent requests per account. The retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted and served by 1017 at the Flash price.

Off-peak rates are half of peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Every other hour, including weekends, bills off-peak. Morph's list rate does not change by time of day.

Read it line by line. On uncached input Morph is below DeepSeek's rate at every hour. On output it is between DeepSeek's off-peak and peak rates. On cache hits it is above DeepSeek's rate at every hour. Which one is cheaper for you depends on your cache-hit rate and when your traffic runs. Many third-party hosts also serve V4.1 Flash; the DeepSeek V4.1 Flash page compares them and works through cache-hit economics.

Running V4.1 on Morph

• Model IDs. morph-dsv41flash and morph-dsv41flash-1017 route to the same deployment at the same price. The older morph-dsv4flash ID still works as an alias and now serves the new weights; use the V4.1 IDs in new code. Context is 1M tokens.
• Endpoints. OpenAI-compatible at https://api.morphllm.com/v1, Anthropic Messages at https://api.morphllm.com. One key also covers Kimi K3.1 Pro (morph-kimik31) and GLM-5.3 (morph-glm53-744b).
• Measured speed. Morph's public hourly charts at ultra.morphllm.com show V4.1 Flash at about 215 output tokens per second and about 0.6 s to first token (median of hourly medians, 24 hours ending November 9, 2026). Artificial Analysis lists V4.1 Flash (Max) at 231 tokens per second on its own method, so the two figures are not a head-to-head.
• Ultra plan. $200 a month for unlimited DeepSeek V4.1 Flash and Kimi K3.1 with no token meter, capped at 1,000 subscriptions. Each subscription gets 701,079 uncached input tokens and 76,559 output tokens per minute on V4.1 Flash. See Ultra.
• OpenRouter. Morph is also a provider for deepseek/deepseek-v4.1-flash on OpenRouter.

Set up a coding agent in one command, then pick the V4.1 model:

curl -fsSL https://www.morphllm.com/install.sh | bash && morph install

Claude Code, Codex and OpenAI SDK configs are on the DeepSeek V4.1 Flash page, and the launch details are in DeepSeek V4.1 Flash on the Morph API. Create a key on the API keys page, and compare every model on Morph models and pricing.

DeepSeek V4.1 Limitations

Frequently Asked Questions