← all stories
DeepSeek
DeepSeek publishes open models and its own inference stack. The most recent release is DeepInfer, the inference system the company runs, published on October 21, 2026. The current model line runs through DeepSeek-V4.1-Flash-1017.
Synthesized from 8
Yomimono stories · updated
16h ago
Across August and October 2026, DeepSeek's coverage settles into a pattern: open models under permissive licenses, low per-token prices, and tooling released around them. The model line moves from DeepSeek-V4-Flash-0731, released as an open model under a commercial-use license in early August, to an MIT-licensed official release of the same model on August 7, then to DeepSeek-V4.1-Flash on September 10, and to DeepSeek-V4.1-Flash-1017 on October 17. In between, DeepSeek released DeepSeek V4 Pro 0813 without a formal announcement on August 14.
The releases did not all land the same way. Third-party testing of V4 Flash found it struggling on real-world agent tasks: Composio ran 30 difficult multi-step workflows across four harnesses and only 6 fully succeeded. DeepSeek raised API prices on August 6, 2026, citing unprecedented demand, and the newer models carry tiered pricing of their own, with peak and off-peak rates for V4.1-Flash-1017 and V4.1-Flash.
On the tooling side, DeepSeek Harness reached version 0.2 on October 1, 2026 with desktop apps for Windows and macOS, plugins, and stated plans for stronger sandboxing, a memory function, and smartphone access. The DeepInfer release builds on vLLM with ideas from SGLang and TensorRT LLM, runs DeepSeek-V4.1-Flash, and supports DSpark acceleration, but a KV cache system is not implemented. DeepSeek also published deepseek-recipe to convert requests and outputs between third-party APIs and the official DeepSeek API format.
Key facts
- Inference system release
- DeepSeek published DeepInfer, the inference system its own AI services use, with core functionality packaged for outside deployment. ↗
- DeepInfer stack
- DeepInfer is built on vLLM with ideas from SGLang and TensorRT LLM, runs DeepSeek-V4.1-Flash, and supports DSpark acceleration. ↗
- DeepInfer gap
- A KV cache system is not implemented in the DeepInfer release. ↗
- Latest model release
- DeepSeek-V4.1-Flash-1017 was released on October 17, 2026 under the MIT license as an updated, faster version of DeepSeek-V4.1-Flash. ↗
- Latest model pricing
- DeepSeek-V4.1-Flash-1017 peak pricing is $0.3 per million input tokens and $1.2 per million output tokens, halved off-peak. ↗
- Harness desktop apps
- DeepSeek Harness version 0.2, released October 1, 2026, added desktop apps for Windows and macOS that can display slide materials, web pages, and code diffs. ↗
- Harness roadmap items
- DeepSeek says more Harness updates are planned, including stronger sandboxing, a memory function, and smartphone access. ↗
- V4.1-Flash specifications
- DeepSeek-V4.1-Flash is a mixture-of-experts model with 552 billion total parameters, 8 billion active parameters on input and 16 billion on output, and it supports image input. ↗
- V4 Pro 0813 pricing
- DeepSeek V4 Pro 0813 is priced at $0.435 per million input tokens and $0.87 per million output tokens, with a discounted cache-hit input rate of $0.003625. ↗
- V4 Flash agent testing
- Composio ran 30 difficult multi-step workflows across four harnesses on V4 Flash, with only 6 fully succeeding. ↗
Timeline
Synthesized by Yomimono from the cited Yomimono stories below, each itself
sourced, then editorially reviewed. Every
fact links the story it came from.
Facts
- Noted
- licenses its model under MIT · 2026-09-11
Connections
Structured graph also available as JSON at /public/entities/deepseek.
CC BY 4.0.
All coverage
3d ago
DeepSeek published DeepInfer, the inference system its own AI services use, with core functionality packaged for outside deployment. It is built on vLLM with ideas from SGLang and TensorRT LLM, runs DeepSeek-V4.1-Flash, and supports the DSpark acceleration technology, but a KV cache system is not implemented. DeepSeek also released deepseek-recipe, which converts requests and outputs between third-party APIs and the official DeepSeek API format.
6d ago
DeepSeek released the open model DeepSeek-V4.1-Flash-1017 on October 17, 2026, positioned as an updated, faster version of DeepSeek-V4.1-Flash. On the DeepSWE v1.1 coding agent benchmark it scores above GPT-6 Astra, GPT-6.1 Sol and Claude Fable 5 while keeping cost low, and it recorded the highest scores on multiple tests. Peak pricing is $0.3 per million input tokens and $1.2 per million output tokens, halved off-peak.
Oct 6
DeepSeek released version 0.2 of its coding agent DeepSeek Harness on October 1, 2026, adding desktop apps for Windows and macOS. The desktop build can display slide materials, web pages, and code diffs, and supports plugins such as one for periodic task execution and one for voice input. DeepSeek says more updates are planned, including stronger sandboxing, a memory function, and smartphone access.
Sep 11
DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026. The open model is a mixture-of-experts system with 552 billion total parameters, 8 billion active parameters on input and 16 billion on output, and it supports image input. It beat Claude Opus 5 and GPT-5.6 Sol on multiple benchmark tests. Inference efficiency cut KV cache tokens to 1/3.9 of DeepSeek-V4-Flash, and off-peak API pricing is half the peak rate.
Aug 19
DeepSeek's V4 Flash model, popular for its low cost and high benchmark scores, is facing criticism after third-party tests showed it struggling on real-world agent tasks. Composio ran 30 difficult multi-step workflows across four harnesses, with only 6 fully succeeding. The company also raised API prices on August 6, 2026, citing unprecedented demand.
Aug 14
Chinese AI startup DeepSeek has released its latest AI model, DeepSeek V4 Pro 0813, without a formal announcement. The release was first noticed on social media and message boards, where users and tech commentators highlighted the model's performance and cost efficiency. According to the open-source AI coding agent Cline, the model scores 15.8 percent higher on Terminal Bench, a benchmark that evaluates how accurately AI can execute real system operations and development tasks on a Linux terminal, compared to the April preview model. Cline also noted that DeepSeek V4 Pro 0813 runs at roughly one fifty-seventh the cost of Claude Fable 5 while delivering comparable performance. The model has 1.6 trillion parameters, 49 billion active parameters, and a 1 million token context window. Tech writer @ChrisGPT reported benchmark improvements across Terminal Bench 2.1, CyberGym, and DeepSWE, with scores rising from 72.1 to 87.9 percent, 52.7 to 83.3 percent, and 12.8 to 62.7 percent respectively. Pricing is set at 0.435 dollars per million input tokens and 0.87 dollars per million output tokens, with a discounted cache-hit input rate of 0.003625 dollars.
Aug 9
DeepSeek has released DeepSeek-V4-Flash-0731 as an open model under a commercial-use license. The MoE model has 284 billion total parameters and 13 billion active parameters. Third-party tests by Artificial Analysis rate its intelligence on par with Google's Gemini 3.6 Flash, and it outperforms GLM-5.2 in coding and GPT-5.6 Luna in agent performance.
Aug 7
DeepSeek has released the official version of its AI model DeepSeek-V4-Flash-0731, replacing the earlier preview version. The weights are available under the MIT license, allowing free commercial use. The model is a Mixture-of-Experts type with 284 billion total parameters, activating only 13 billion during inference, and supports a context length of 1 million tokens. According to benchmark scores, it surpasses the preview version of the higher-tier DeepSeek-V4-Pro, which is more than five times larger, across all nine agent-based benchmark items. Among open models, it also outperformed GLM-5.2 from China in all eight items where scores were published, though it does not reach Anthropic's Claude Opus 4.8 in any item. Community quantized versions are available, including GGUF-format versions from Unsloth in 13 steps ranging from 1-bit at 82.5GB to 8-bit at 162GB. For paid cloud usage, DeepSeek's documentation lists the price at $0.14 per million tokens for input, $0.0028 per million tokens for cached input, and $0.28 per million tokens for output, with prices doubling during peak times.