DeepSeek has executed one of the most aggressive model release cycles in artificial intelligence history. Between April and August 2026, the Hangzhou-based lab shipped two model families, released 893 gigabytes of open weights under an MIT license, introduced the first major AI provider to implement surge pricing, launched an open-source agent harness that attracted 95,000 GitHub stars in two days, and released an experimental multimodal model that claims to rival Anthropic's Opus 4.8 on agent benchmarks.
This article covers every major release in the DeepSeek V4 era: the architecture, the benchmarks, the pricing, the open weights, the agent harness, and the multimodal vision model. All claims are traced to primary sources including DeepSeek's own API documentation, technical papers on arXiv, Hugging Face model cards, and independent reporting.
From V3 to V4: The Evolution
DeepSeek V3, released in December 2024, established the lab's reputation for efficiency. It was a Mixture-of-Experts model with 671 billion total parameters and 37 billion activated per token, trained on 14.8 trillion tokens at a total cost of 2.788 million H800 GPU hours, which the authors calculated at approximately $5.576 million. The training used FP8 mixed precision, a technique the paper describes as the first validation of FP8 training at this scale. V3 adopted Multi-head Latent Attention (MLA) for efficient inference and DeepSeekMoE for cost-effective training. It also introduced an auxiliary-loss-free strategy for load balancing and a Multi-Token Prediction (MTP) objective that improved both training efficiency and enabled speculative decoding for inference acceleration.
The V3 chat version outperformed other open-source models and achieved performance comparable to GPT-4o and Claude-3.5-Sonnet on standard benchmarks, according to the paper published in Nature.
DeepSeek V4, previewed on April 24, 2026, represented a fundamental architectural shift. The lab moved away from MLA to a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), as detailed in the V4 technical paper on arXiv. It replaced conventional residual connections with Manifold-Constrained Hyper-Connections (mHC), which constrain the residual mapping onto a specific manifold to enhance signal propagation stability across layers. The optimizer changed from AdamW to Muon, which the paper credits with faster convergence and greater training stability. Both V4 variants were pre-trained on more than 32 trillion tokens, more than double V3's dataset.
The two V4 variants are V4-Pro, with 1.6 trillion total parameters and 49 billion activated per token, and V4-Flash, with 284 billion total parameters and 13 billion activated per token. Both support a one-million-token context window with a 384K maximum output length.
V4-Flash-0731: The Budget Agent
V4-Flash-0731 entered public beta on July 31, 2026. It is a re-post-trained variant on the unchanged architecture from the April preview. Its benchmark results on DeepSeek's own harness include Terminal Bench 2.1 at 82.7, NL2Repo at 54.2, Cybergym at 76.7, DeepSWE at 54.4, Toolathlon-Verified at 70.3, DSBench-FullStack at 68.7, and DSBench-Hard at 59.6.
Artificial Analysis measured V4-Flash as the cheapest well-known AI model to run, completing its Intelligence Index test battery at approximately three cents. The nearest comparisons were Moonshot's Kimi K3 at 86 cents, OpenAI's GPT-5.6 Sol at $1.86, and Anthropic's Claude Fable 5 at $3.15. V4-Flash scored 50 on the Intelligence Index, level with Google's Gemini 3.6 Flash.
The official API pricing for V4-Flash off-peak is $0.22 per million input tokens and $0.66 per million output tokens, as listed on the DeepSeek pricing page. Peak rates double to $0.44 and $1.32 respectively. Cache-hit input costs just $0.0028 per million tokens.
V4-Pro-0813 GA: The Flagship
DeepSeek-V4-Pro went to general availability on August 13, 2026, announced via a changelog entry on the API docs. The release brought significantly enhanced agent capabilities, native OpenAI Responses API support adapted for Codex, and three thinking effort levels (low, high, max) for both V4-Pro and V4-Flash.
The GA benchmark table published by DeepSeek shows V4-Pro-0813 scoring HLE with tools at 60.0, Terminal Bench 2.1 at 87.9, NL2Repo at 61.5, Cybergym at 83.3, DeepSWE at 62.7, Toolathlon-Verified at 74.1, Agents' Last Exam at 25.7, AutomationBench at 31.8, DSBench-FullStack at 71.1, and DSBench-Hard at 67.2.
The release pattern was unusual. The version string DeepSeek-V4-Pro-0813 appeared in the price list on August 12 with no announcement. The changelog entry dated August 13 followed the next day. The benchmark table reportedly leaked via a WeChat group before landing on Hacker News, according to Simon Willison's blog. There is no press page for the release. The MIT-licensed weights, approximately 893 gigabytes across 67 safetensors files, appeared on Hugging Face the same day as the changelog entry.
The official API pricing for V4-Pro off-peak is $0.66 per million input tokens and $1.98 per million output tokens, with peak rates of $1.32 and $3.96 as documented by DeepSeek. Cache-hit input costs $0.0036 per million tokens. The V4-Pro-0813 weights include a speculative decoding module called DSpark.
Surge Pricing: The First Major AI Provider to Do It
DeepSeek announced on August 13 that it was introducing peak and off-peak pricing, effective at 16:00 UTC on August 16. This makes DeepSeek the first major AI provider to implement structural surge pricing for model inference.
Peak hours are defined as 01:00 to 04:00 UTC and 06:00 to 10:00 UTC. Off-peak rates are set at 50% of peak-hour prices. DeepSeek's stated rationale is to manage immense global demand for compute cycles by incentivizing developers to schedule batch processing and non-urgent workloads during lower-demand windows. The model treats compute like electricity, with rates varying by time of day.
This pricing change came after DeepSeek had made a 75% discount permanent earlier in 2026. The company warned of a "significant increase" in API pricing, which has made the open-weights route more attractive for organizations that can self-host.
Open Weights: 893 GB on Hugging Face
The V4-Pro-0813 weights are available on Hugging Face under an MIT license. The repository deepseek-ai/DeepSeek-V4-Pro-0813 contains 67 safetensors files totaling approximately 893 gigabytes. The V4-Flash-0731 variant is similarly available at deepseek-ai/DeepSeek-V4-Flash-0731.
The weights are open in the licensing sense, not the laptop sense. Running V4-Pro requires serious hardware: expert-parallel deployment across multiple GPUs, with each GPU hosting only one expert. The minimum deployment unit for the decoding stage consists of 40 nodes with 320 GPUs, according to the V3 paper's deployment section, which applies to the V4 architecture with its larger expert count.
For organizations that can handle the infrastructure, the MIT license permits download, run, and fine-tune without restriction. For teams that cannot self-host, the API remains the reference deployment at the pricing rates described above.
The V4 paper describes several architectural innovations over V3. The hybrid attention architecture uses Compressed Sparse Attention (CSA), which compresses the KV cache of every token into one entry and then applies DeepSeek Sparse Attention where each query token attends only to compressed KV entries, and Heavily Compressed Attention (HCA), which consolidates the KV cache of every 128 tokens into a single entry with dense attention. The two are interleaved across layers.
Manifold-Constrained Hyper-Connections replace residual connections, constraining the residual mapping onto a specific manifold to enhance stability of signal propagation across layers. The Muon optimizer replaces AdamW for faster convergence. The MoE activation function changed from Sigmoid to Sqrt(Softplus). Hash routing is used in the initial Transformer blocks. Multi-Token Prediction modules are retained from V3. FP4 quantization-aware training is applied to MoE expert weights and the QK path in CSA's indexer, achieving a 2x speedup for the top-k selector while preserving a 99.7% recall rate of KV entries.
DeepSeek Harness: Everything Is a Plugin
DeepSeek Harness (dsh) was released on August 13, 2026, alongside the V4-Pro GA. It is a TypeScript-based agent harness distributed under the MIT license. The project's tagline is "Everything is a plugin."
The framework passed 95,386 GitHub stars and 8,826 forks within approximately two days of publication, according to the GitHub API. It serves a local web UI on port 3080 via the command npx @deepseek-ai/dsh web.
The architecture treats every capability as a Cordis plugin: the model adapter, the tool registry, the session log, the sandbox, and the agent loop itself. There is no privileged core to patch. The framework is model-agnostic, supporting DeepSeek, Anthropic, OpenAI, and custom OpenAI-compatible endpoints.
Researchers from Peking University and DeepSeek-AI published an 88-page paper formalizing the plugin architecture, which proves that plugin removal and dependency changes are safe at the system level and validates the design on a production ecosystem of more than 4,000 plugins built over four years.
The benchmarks for V4-Pro-0813 were tested using the Harness in minimal mode. For public code-agent tasks, DeepSeek states this explicitly in its benchmark documentation.
DeepSeek Harness competes with Claude Code (Anthropic's commercial terminal agent) and OpenAI Codex (OpenAI's commercial IDE agent), both of which are model-locked to their respective vendors. Claude Code supports Claude models primarily, with extension points for skills, hooks, MCP, and plugins. OpenAI Codex is optimized for OpenAI models with skills, MCP, and custom agents. dsh's distinguishing bet is that every component is replaceable, including the model itself.
V4-Flash-Vision-Exp: Multimodal Agents
On August 21, 2026, DeepSeek released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal variant that adds image understanding to the text-only V4-Flash model. The release was announced on the DeepSeek API platform and is available via model='deepseek-v4-flash-vision-exp'.
The model matches V4-Flash on text capabilities including agents, reasoning, and world knowledge. On multimodal agent benchmarks, DeepSeek claims it brings performance close to Anthropic's Claude Opus 4.8.
DeepSeek published eleven benchmark results comparing Vision-Exp against Opus-4.8. The new model beats Opus-4.8 on three benchmarks: DeepSWE (59.3 vs 58.0, a margin of 1.3 points), Agents' Last Exam (27.3 vs 25.7, a margin of 1.6 points), and ZeroBench (35.0 vs 34.0, a margin of 1.0 point). It trails on the other eight, with the largest gap on NL2Repo (57.7 vs 69.7, a difference of 12 points) and DSBench-Hard (63.6 vs 71.7).
The close results are genuinely close. Toolathlon-Verified splits 75.9 to 76.2. Chartography splits 64.3 to 65.0. Terminal Bench 2.1 is 83.9 against 85.0.
The headline claim is the jump in multimodal agent performance over V4-Flash. On ApexBench, the vision variant scores 36.5 against V4-Flash's 26.2. On Agents' Last Exam, it scores 27.3 against 25.2. DeepSeek's own footnote explains that in these evaluations, the text-based V4-Flash "ignores multimodal elements contained therein," meaning the older model was being scored on tests containing images it cannot see.
The vision model supports JPEG, PNG, GIF, and WebP files. Images are tokenized for billing at up to 384 tokens each, at V4-Flash pricing. A single request can carry up to 600 images, with a maximum edge length of 8,192 pixels (4,096 pixels once a request holds 15 or more images). The model normalizes any image to approximately 800 by 800 pixels before processing. An optional "detail" setting downscales images to 512 by 512 pixels to save tokens.
Images can be provided via Base64, external URLs up to 32 MiB, or the new Files API, which is free to use with a 64 MiB limit and allows one upload to be referenced by file_id across many requests.
Vision-Exp works across OpenAI's Chat Completions and Responses APIs, Anthropic's Messages endpoint, and DeepSeek Harness 0.1.1, which was released the same day with out-of-the-box support.
DeepSeek has not said when or whether Vision-Exp graduates into a production model the way V4-Flash-0731 did. Teams evaluating multi-vendor agent stacks should treat the API compatibility, not the benchmark score, as the reason to test this model: it is the cheapest way right now to confirm an agent pipeline still works if a primary model provider changes.
Market Impact and the Price War
DeepSeek's pricing strategy has consistently disrupted the market. The company made a 75% discount permanent earlier in 2026, and rivals have been cutting in response. OpenAI trimmed GPT-5.6 pricing sharply. The broader drift of the market has been down, fast, on a curve that looks less like software margins and more like a commodity.
The company recently closed its first outside funding round of more than $7 billion, which buys room to subsidize aggressive pricing while it takes share. Zack Kass, OpenAI's former head of go-to-market, has framed the moment as one of "diminishing model returns," arguing that once models are close enough, the next one barely moves the needle and price does the deciding.
The competition is no longer just about who builds the best text model. It is about who builds the most capable and affordable multimodal system, who controls the agent harness layer, and who offers the most flexible pricing model. DeepSeek is now competing on all three fronts simultaneously.
What This Means
DeepSeek has built a complete stack: two production-grade models (V4-Pro and V4-Flash), an experimental multimodal variant (Vision-Exp), an open-source agent harness (dsh), open weights under MIT, and a pricing model that treats compute like electricity. The lab is no longer competing only on model intelligence and token prices. With Harness, it moves into the layer that determines how models use tools, manipulate files, maintain sessions, and execute long-running workflows, which is exactly the territory where Claude Code became a business.
The benchmark numbers are vendor-run on DeepSeek's own harness, and no independent replication has been published at the time of writing. The claims should be read as vendor claims until a third party reruns them. The open-weights route is real, but running V4-Pro requires serious infrastructure. The pricing is the cheapest in the industry by independent measurement, but surge pricing adds complexity to cost modeling.
For teams building visual agents on a budget, DeepSeek's latest offering is now one of the most compelling options on the market. For teams building agentic workflows, the combination of open weights, a model-agnostic harness, and multi-vendor API compatibility makes DeepSeek's stack uniquely flexible.
For more on the broader AI landscape, see Nvidia's Groq deal for agentic AI inference acceleration and Nvidia pays Poolside $6 billion for AI model licensing.