DeepSeek released V4.1-Flash on September 10, 2026, an open-weight model with a causal encoder-decoder architecture, native vision, and a 1 million token context window, under an MIT license. The company also published a technical report. The model carries 763B total parameters, with 8B active parameters for prefill and 16B active for decode.
The release is positioned as a major architectural overhaul despite the modest version number. DeepSeek V4 Pro, released in April 2026 alongside Engram, is being soft-retired. Traffic to the older model is being routed to V4.1-Flash at cheaper pricing until V4.1 Pro launches. Deprecated models deepseek-v4-flash, deepseek-v4-flash-vision-exp, and eventually deepseek-v4-pro will route to V4.1 Flash.
DeepSeek's own announcement tweet drew 6.04M views, 933 replies, 3.02K reposts, and 27.7K likes. A second DeepSeek tweet drew 176K views, 56 replies, 143 reposts, and 1.67K likes. The announcement tweet carried the CSS class string "pencraft pc-display-flex pc-flexDirection-column pc-gap-12 pc-padding-16 pc-reset bg-primary-zk6FDl outline-detail-vcQLyr pc-borderRadius-md sizing-border-box-DggLA4 pressable-lg-kV7yq8 font-text-qe4AeH tweet-fWkQfo twitter-embed" in its embed markup, a detail noted by observers tracking the release.
Architecture and Efficiency
The model uses a novel causal Encoder-Decoder design. It supports native visual understanding, meaning text and image input. Context length is 1 million tokens. The license is MIT.
The architecture includes Sliding-Window Attention Bounded Replay. The model has 40 layers, arguably only 20 legit decoder layers. It uses 3x3 pixel unshuffle for vision instead of the common 2x2.
KV cache footprint is up to 1/8 that of V4 Flash. KV cache size is about 890 bytes per token in the benchmarked score regime. KV-cache and storage reductions are 4x HBM and 8x SSD versus the prior generation, and 437x versus the first-generation model.
Parameter count confusion emerged after release. Hugging Face safetensors show about 748.5B parameters for backbone plus engram: 551.566B backbone plus 196.929B engram. Including DSpark/MTP at 14.225B and the vision encoder at 0.485B brings the total to about 763.21B parameters, or 511.76 GB. Backbone MoE FFN experts account for 543.582B parameters in FP4. Attention, shared, embedding, and other components account for about 7.984B. Some claim only about 9B active parameters for prefilling.
A Reddit post titled "Deepseek V4.1 Flash is 748B, not 552B" drew 575 activity. Another post, "Deepseek Has Soft Retired Deepseek V4 Pro," drew 1598 activity. A third, "DeepSeek V4.1 Flash: Stronger, Faster, More Accessible," drew 317 activity.
Sebastian Raschka, an AI researcher and educator, publicly commented that V4.1 should have been called V5 due to the architectural overhaul. His tweet drew 6.72K views, 8 replies, 3 reposts, and 120 likes. Peter Gostev, an AI commentator, created a 3D visualization comparing V4.1-Flash to the original Transformer. His tweet drew 218K views, 54 replies, 334 reposts, and 2.76K likes.
Stochastic Chasm, a technical commentator, compared the architecture to prior work and noted multimodal details. Stochastic Chasm said the model "lets the backbone handle most of it and give it visual tokens," and described the layer count as "down to 40 layers , arguably only 20 legit decoder layers ," and summarized the efficiency story as "how obsessing over KV Cache compression gets you a hyper-efficient frontier model."
Stochastic Chasm also called the design "a very interesting mix of very conservative and sometimes old ideas in research and potentially cutting edge efficiency and hardware design in engineering."
Nrehiew, a technical commentator, highlighted KV cache compression and infrastructure details. Nrehiew framed the report as "cleaner" than the prior HSA/CSA design and likely unlike what OpenAI or Anthropic would do because of their custom chips. That is informed opinion.
Benchmarks and Pricing
Artificial Analysis, an independent benchmark organization, reported benchmark scores and pricing for V4.1-Flash. The model scored 40 on the Artificial Analysis Intelligence Index. It scored 69% on AutomationBench-AA, tying GPT-6 Astra and above Grok 4.6 at 67%. On GDPval-AA v2 it scored 1632 Elo, up from 1468, overtaking Kimi K3 at 1584. On AA-LCR v1.1 it scored 84%, on par with GPT-5.6 Sol and Gemini 3.8 Flash.
Average verbosity was 89k tokens per Intelligence Index task, 25% more than GLM-5.3 at 71k, 29% more than GLM-5.3-Flash at 69k, and 62% more than V4 Pro 0813 at 55k. That is above Fable 5.1 at 78k and Claude Opus 5 at 73k.
Estimated cost per Intelligence Index task was $0.27, about 7x below GLM-5.3 at $2.01 and Kimi K3 at $2.00, and about 2.5x below V4 Pro 0813 at $0.67.
Vals, an AI benchmark and index provider, ranked V4.1-Flash as the number one open-weight model on its index, ahead of Kimi K3. On the Vals Index the model costs $0.30 per test, the cheapest in the open-weight top 10. Vals eval settings were 1M context, 384 max output tokens, temperature 1, default top-p and top-k, and high reasoning effort.
Artificial Analysis said V4.1-Flash surpasses DeepSeek V4 Pro 0813 despite being much cheaper. It also said the model is among the most verbose models measured, but that even with verbosity, cost per task is exceptionally low.
TeortaxesTex, a technical commentator, called some eval results "very strange," particularly AutomationBench number one and a CritPt regression. TeortaxesTex also argued DeepSeek often shows high internal evals, weaker external robustness, brittleness, and weird skill gaps. TeortaxesTex argued V4 GA benefited massively from tool and skills harness access, whereas V4.1 appears less dependent on harness scaffolding. TeortaxesTex reported multi-agent "DSH agent teams" could degrade quality unless a project has very clear modularity.
TeortaxesTex speculated that DeepSeek "repeatedly struggled to train first layers properly." That is inference, not confirmed by DeepSeek.
Jared Z, an AI commentator, critiqued that users care deeply about token cost, and daily-driver coding models should be both cheap and smart.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Local Inference and Deployment
Baseten, a model deployment platform, shipped day-0 support for V4.1-Flash. Ollama, a local LLM runner, rolled out V4.1-Flash support to Max, Team, and later Pro subscribers.
Fraser Price, an inference engineer, reported running V4.1-Flash locally at high speeds with offloading. Local inference hit 200 TPS on 4 Max-Qs with 64GB system RAM, offloading a 200GB Engram and hash table to NVMe. It later reached 300+ TPS on 4 RTX Pros with less than 32GB peak system RAM.
Antirez, the developer Salvatore Sanfilippo, demonstrated V4.1-Flash running on a 128GB M5 Max with SSD streaming. Fraser Price called it "incredible you can run frontier models mostly off SSD."
A vLLM release included DeepSeek-V4 shared experts fused into MegaMoE, and Mooncake Store can offload decode KV.
Post-Training and Research Direction
DeepSeek's post-training notes say the company agrees the ROI of improving data quality far exceeds novel post-training algorithms. Jasper Lu, an AI commentator, noted DeepSeek's post-training comments on data quality ROI.
Shikib Mehri, an AI researcher, pushed back on "research is over" claims. Mehri argued the lever surface expanded from architecture into data-factory and reward-design research. Assertions that "data is all that matters" or "research is over" were themselves criticized as overreactions.
DeepSeek remains polarizing. It is admired for shipping unusual research artifacts and detailed reports, but seen by some as less polished than product-centric competitors. One external judgment was that "DeepSeek ships internal research artifacts and not products."
Commenters speculated V4 Pro's GA release may have suffered from reward hacking and poor scaling. Some argued V4 Pro showed "high degree of reward hacking" and was "not performing meaningfully better than the flash model despite being nearly 6 times the size." Commenters speculated smaller "Flash" variants outperforming larger models may not be simple distillations from one large training run. Some users distinguished between API retirement and model disappearance: DeepSeek stopped serving V4 Pro, but weights reportedly remain available. A technical hypothesis was that DeepSeek may be freeing inference capacity or migrating toward Chinese inference chips.
DeepSeek's research agenda includes intermediate papers with hyperfocused architectural improvements and a high hit rate, from Math (GRPO), Coder, R1, to Manifold Constrained Hyperconnections and Compressed Sparse Attention. After the R1 paper in 1H2025, DeepSeek laid low for about a year, letting peers like GLM and Kimi take the lead on open models. DeepSeek V4 Pro and Engram were released in April 2026.
The launch lands amid a broader shift from bigger dense chat models toward systems-optimized, sparse, long-context, agent-oriented models. The architecture reflects a broader industry trend toward splitting prefill and decode economics. The release reinforces the idea that open models are increasingly competitive on servability. It also sharpened debate over what matters most in 2026 model progress: architecture, RL and inference co-design, data quality, or systems work.
The Wider Release Week
OpenAI launched GPT-Live-1 into the API, a full-duplex voice interface that can listen while speaking and delegate tool use and reasoning. GPT-Live-1 scored 83.6% first-attempt task completion on Tau3 with GPT-6 Astra, 97.3% on Artificial Analysis Conversational Dynamics, and 0.798s response onset latency on Full Duplex Bench v1. OpenAI announced a public-beta Agents API with the Codex harness, and hosted sandboxes for code execution, files, and artifacts. OpenAI paused new $200 Pro signups for Astra capacity reasons.
Integration announcements from LiveKit, HeyGen, Telnyx, Speak, and Cognition's Devin Voice suggest GPT-Live-1 may become a default substrate for production voice agents. OpenAI's product-side announcement of a Data agent in ChatGPT Work promises dashboards, answers, and actions over connected company data sources. Box framed its integration as "the file system for AI," bringing governed enterprise context into ChatGPT.
Cognition released SWE-2, described as "our closest model yet to the frontier," claiming parity on leading coding evals at up to 70% lower cost, and scaled RL to multiple trillions of parameters. Cognition launched Devin Voice powered by GPT-Live and SWE-2. Dioxus Labs is joining Cognition to contribute to Devin's VM, computer use, and testing.
Cursor introduced Projects: persistent threads with a coordinator agent, shared memory and artifacts, and sync across devices. Claude Code added new pane pop-outs and a managed-agent session viewer and auto mode. Google's docs-for-agents push and Cursor's new persistent workspaces trend toward stateful, organization-aware agent environments.
Several papers pushed on the theme that the harness is now a core optimization target. A Salesforce paper found that training a weaker model on a stronger expert's full trajectories can hurt performance by 4 to 30 points after harness evolution. The fix is to rewrite only the failing turn in the weaker model's own rollout. ByteDance's HarnessDev described agents that build and improve their own runnable harnesses, with only 34 of 64 changes transferring directionally to held-out tasks. Qwen's Elastic Horizon is a closed-loop controller that tracks the 90th percentile of successful trajectory lengths, improving success while saving up to 25% of trajectory tokens. PARSER replaces sequential chunk reading with parallel frozen subagents and an RL-trained lead agent, reporting +12 points at 896K context and up to 11x lower latency. SkillAdam framed skill self-evolution as a discrete optimization problem using Adam-like first and second moment ideas. Google Research's ToolGrad generates ground-truth tool-use chains before prompts, with near-100% pass rate for dataset creation and downstream tool-use gains.
Anthropic published its most detailed misuse report, covering attempts to use Claude for cyberattacks, influence ops, surveillance, biology, and weapons, and disrupted every operation described. Redwood Research proposed transparency norms for architectures that may weaken or eliminate chain-of-thought visibility. Ryan Greenblatt argued companies should publish evidence and policies before deploying architectures that substantially reduce CoT dependence. Neel Nanda interpreted GPT-6 Astra as a potentially concerning jump in no-CoT reasoning. Disagreement among frontier-lab employees and alumni about risk culture continued. Thom Wolf announced a new Open Alignment team at Hugging Face. Richard Ngo published a critique of Paul joining OpenAI's board and of the safety community's capture by AGI companies.
The release period covered by AINews was September 9-10, 2026. DeepSeek's V4.1-Flash announcement came on September 10, 2026. Peter Gostev posted his 3D visualization the same day. Sebastian Raschka posted his meme about V4.1 being V5 on September 11, 2026. One observer summarized the overall impression of the release as an "interesting release / breath of fresh air vibe," capturing the tone among some commenters.