# DeepSeek: what I've logged

> DeepSeek comes up in 11 posts of my AI log, from September 2025 to September 2026, and each item below links to the primary source. The releases I logged include DeepSeek V4.1 Flash, DeepSeek V4 Flash Vision Exp and DeepSeek-V4 Preview, filed under open weights, agents and coding and science.

## Items

1. **DeepSeek V4.1 Flash** (DeepSeek, 7 Sept 2026): A 552B mixture-of-experts model under MIT with native vision, 1M context and a new encoder-decoder design that uses 8B active parameters for input and 16B for output. [Primary source](https://www.deepseek.com/en/news/deepseek-v4-1-flash/) · [Post](https://heyaliasif.com/ai/2026-09-07-week-in-ai#item-1)
2. **DeepSeek V4 Flash Vision Exp** (DeepSeek, 17 Aug 2026): An experimental image-capable version of V4 Flash that matches its text quality, billed at V4 Flash text rates with images costing up to 384 tokens each. [Primary source](https://api-docs.deepseek.com/news/news260821) · [Post](https://heyaliasif.com/ai/2026-08-17-week-in-ai#item-2)
3. **DeepSeek-V4 Preview** (DeepSeek, 20 Apr 2026): Open weights for V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active), both with a 1M-token default context. [Primary source](https://api-docs.deepseek.com/news/news260424/) · [Post](https://heyaliasif.com/ai/2026-04-20-week-in-ai#item-2)
4. **DeepSeek-OCR 2** (DeepSeek, 26 Jan 2026): A second version of DeepSeek's open document reader, which matters to anyone digitising paper-heavy workflows. [Primary source](https://huggingface.co/deepseek-ai/DeepSeek-OCR-2) · [Post](https://heyaliasif.com/ai/2026-01-26-week-in-ai#item-3)
5. **Conditional Memory via Scalable Lookup (Engram)** (DeepSeek, 12 Jan 2026): DeepSeek adds a lookup-based memory as a second axis of sparsity next to mixture-of-experts. [Primary source](https://arxiv.org/abs/2601.07372) · [Post](https://heyaliasif.com/ai/2026-01-12-week-in-ai#item-6)
6. **mHC: Manifold-Constrained Hyper-Connections** (DeepSeek, 29 Dec 2025): DeepSeek proposes a way to widen residual connections in transformers while keeping training stable, a possible architecture lever for future models. [Primary source](https://arxiv.org/abs/2512.24880) · [Post](https://heyaliasif.com/ai/2025-12-29-week-in-ai#item-4)
7. **DeepSeek-V3.2** (DeepSeek, 1 Dec 2025): Open weights with sparse attention and tool use built into the reasoning; a higher-compute Speciale variant targets hard problems. [Primary source](https://huggingface.co/deepseek-ai/DeepSeek-V3.2) · [Post](https://heyaliasif.com/ai/2025-12-01-week-in-ai#item-1)
8. **DeepSeekMath-V2** (DeepSeek, 24 Nov 2025): An open math model trained to verify its own proofs, aimed at theorem proving. [Primary source](https://huggingface.co/deepseek-ai/DeepSeek-Math-V2) · [Post](https://heyaliasif.com/ai/2025-11-24-week-in-ai#item-5)
9. **DeepSeek-OCR** (DeepSeek, 20 Oct 2025): Tests compressing a page of text into a small number of vision tokens, a possible route to cheaper long context. [Primary source](https://huggingface.co/deepseek-ai/DeepSeek-OCR) · [Post](https://heyaliasif.com/ai/2025-10-20-week-in-ai#item-3)
10. **DeepSeek-V3.2-Exp** (DeepSeek, 29 Sept 2025): Introduces sparse attention to cut long-context cost, and DeepSeek halved its API prices alongside it. [Primary source](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp) · [Post](https://heyaliasif.com/ai/2025-09-29-week-in-ai#item-4)
11. **DeepSeek-V3.1-Terminus** (DeepSeek, 22 Sept 2025): A refinement of V3.1 that fixes mixed-language output and improves its code and search agents. [Primary source](https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Terminus) · [Post](https://heyaliasif.com/ai/2025-09-22-week-in-ai#item-3)

## Posts

- 2026-09-07 [DeepSeek V4.1 Flash, Devin's SWE-2 and a petabyte of DNA predictions](https://heyaliasif.com/ai/2026-09-07-week-in-ai)
- 2026-08-17 [A quiet week: Qwen3.8-27B, DeepSeek adds vision, DeepMind on games](https://heyaliasif.com/ai/2026-08-17-week-in-ai)
- 2026-04-20 [GPT-5.5, DeepSeek V4 preview and Kimi K2.6 in one week](https://heyaliasif.com/ai/2026-04-20-week-in-ai)
- 2026-01-26 [Kimi K2.5, DeepMind's Project Genie, and the OpenClaw frenzy](https://heyaliasif.com/ai/2026-01-26-week-in-ai)
- 2026-01-12 [Anthropic's Cowork, DeepSeek's Engram and a new open image editor](https://heyaliasif.com/ai/2026-01-12-week-in-ai)
- 2025-12-29 [DeepSeek's mHC paper, Qwen-Image-2512 and a Tencent translation model](https://heyaliasif.com/ai/2025-12-29-week-in-ai)
- 2025-12-01 [DeepSeek V3.2, Mistral 3, and Anthropic buys Bun](https://heyaliasif.com/ai/2025-12-01-week-in-ai)
- 2025-11-24 [Claude Opus 4.5, FLUX.2, and a wave of strong open releases](https://heyaliasif.com/ai/2025-11-24-week-in-ai)
- 2025-10-20 [Claude Code on the web, DeepSeek-OCR, Anthropic's TPU deal](https://heyaliasif.com/ai/2025-10-20-week-in-ai)
- 2025-09-29 [Claude Sonnet 4.5, Sora 2, and DeepSeek's sparse attention](https://heyaliasif.com/ai/2025-09-29-week-in-ai)
- 2025-09-22 [Qwen3-Omni and VL arrive, Gemini Robotics 1.5 learns to think first](https://heyaliasif.com/ai/2025-09-22-week-in-ai)

Canonical: https://heyaliasif.com/ai/topics/deepseek