DeepSeek News

DeepSeek News

Follow the latest DeepSeek model releases, API changes, pricing and developer tools, alongside news from the wider AI ecosystem. DeepSeek V4 Hub is an independent publication.

Latest news

135
Analysis

Qwen 4-Max-Preview Is Live Before the Full Launch. What the Early Tests Show (Nov 2026)

Alibaba announced Qwen 4 on November 11, 2026, but the full model is not out. Qwen 4-Max-Preview is live in QwenWork, Qoder, Qwen App and QwenStudio, with no release date for the stable version. The official pitch is deeper knowledge, richer world understanding, and stronger coding and agentic capabilities, aimed at complex coding, knowledge work, front-end builds and professional software workflows. Early public evidence is mixed: three working SVG animations with rough visuals, and a GameFeel RE score of 10.3 that trails Qwen3.8 Flash and Max. Qoder also says the preview changed daily during four anonymous test days, so every review is tied to a build.

Read article
Analysis

TCSExam: What GPT-6 Astra and Kimi K3.1 Pro Can Actually Prove (Nov 2026)

TCSExam is a 100-task benchmark for theoretical computer science research, co-authored by Moonshot AI's Zhilin Yang and Turing Award winner Andrew Yao. Every task ends in an artifact that Lean or an executable checker can verify: a proof, a solver, a construction, or an answer to an open conjecture. GPT-6 Astra leads the Full set at 71.9 and Kimi K3.1 Pro is second at 57.4. The useful part is the spread: Astra solves 70.0% of the bare theorem tasks, Kimi 38.3%, and the other six agents stay at or below 11.7%. On the smaller Lite set, Opus 5.5 leads at 72.3.

Read article
Analysis

Inside HarnessPro: The Prompt Rules That Keep a Self-Evolving Agent Honest (Nov 2026)

HarnessPro freezes the model and lets a search loop rewrite the harness around it: instructions, tools, hooks and budgets. The reusable part is the prompt contract. A test-writing agent replaced an LLM judge that caught only 41.7% of failures, and its frozen tests caught 83.3%. Three rules keep the search from grading itself into a better score. The 53% saving is a deployment cost per task on SWE-Bench Pro, not the cost of running the evolution.

Read article
Analysis

DSH Go Put DeepSeek Harness on Android. The 20,000-Star Claim Is a Week Off

DSH Go is a solo-built, open-source Android client for DeepSeek Harness. It runs the harness inside a bundled Ubuntu environment on the phone, or connects to one already running on a desktop over SSH or WebRTC. On November 11, 2026 the repository had 18,532 stars, 814 forks and 469 open issues plus 32 open pull requests. The viral "nearly 20,000 stars in 10 days" version is about a week off: the project had roughly 8,400 stars at day 10 and would reach 20,000 around day 18 at its recent pace. Combined with the same author's earlier dsh-TUI plugin, the two projects now hold about 25,900 stars.

Read article
Analysis

Anthropic's Agent Memory Architecture: Four Layers, Dreaming and a 413-File Failure (Nov 2026)

Anthropic published The components of agent memory on November 6, 2026, describing the memory stack it runs in Claude Code: raw transcripts, memory notes, a linked wiki, and a small index injected at session start. The counterintuitive part is the failure record. In Pokemon Red, Sonnet 3.7 wrote 413 files across 225 folders by about step 10,000 and won one badge, while Fable 5 kept nine files in two folders plus a global.md and won four. A later offline pass called dreaming caught a wrong note that had sent three runs over the same one-way ledge.

Read article
Analysis

Kimi K3.1 Pro Highspeed vs Claude Fable 5.2: Which Is Better for Design and Coding?

Kimi K3.1 Pro Highspeed costs half as much per uncached input or output token as Claude Fable 5.2 at their listed API rates. Start with Kimi if you want to iterate on designs while keeping costs down. Test Fable on harder coding or research work where the quality of the result matters more than the token price. For aesthetics, we have not found a direct comparison of these exact versions, so judge them against your own briefs.

Read article
Analysis

DeepSeek Harness is quietly winning Latin America: $16 for 2 billion tokens, and an old Motorola phone running 47 websites (Nov 2026)

DeepSeek Harness paired with the official DeepSeek API has become the default budget stack across Latin America's Spanish-speaking developer communities. Real bills from the region: 1.985 billion tokens in one month for $14.25 plus ¥3.89, and 3.458 billion tokens over 30 days for about $30.50, helped by cache hit rates above 99 percent and an off-peak window that covers the Latin American workday. The measured catch: ETH Zurich's MathArena logged the same DeepSeek-V4.1-Flash-1017 (Max) model at 51.6% expected score in its reference setup versus 43.4% under DeepSeek Harness, with cost per problem down from $1.15 to $0.35.

Read article
Analysis

Is DeepSeek V4.1 Pro in Testing? What the 64% LobeBench Result Shows

A LobeHub evaluator reports that the current DeepSeek V4 Pro on the web completed 32 of 50 difficult agent tasks in all three attempts, up from 24 for V4 Pro-0813. The result has sparked speculation about V4.1 Pro, but the test does not establish the identity of the model behind the service. DeepSeek has not announced a V4.1 Pro release date.

Read article
Analysis

Kimi K3.1 Pro vs GPT-6 Astra: Four Blender Render Tests and a Quota Limit

In four Blender tasks shown by Xiyangshi AI Video, Kimi K3.1 Pro and GPT-6 Astra produced comparable-looking renders from the same prompts, with different details and repairs. The creator reported hitting a five-hour Kimi usage limit during the session. These are illustrative third-party demonstrations, not an independently reproduced benchmark or a general ranking. Separately, Artificial Analysis scored Kimi K3.1 Pro 55 in its October 23, 2026 launch snapshot.

Read article