ai-powered-markdown-translatorArticle translated from fr to en with gpt-5.4-mini.
August 3, 2026 brings three major announcements: Alibaba introduces Qwen3.8-Max, its new flagship 2.4T-parameter model; Genspark opens up the source code of its GenOffice productivity suite; and Vercel opens a programmable API for v0. Around these three releases, Cursor, GitHub, Sakana AI, MiniMax, Runway, ElevenLabs, and Google are multiplying product updates.
Qwen3.8-Max, Alibaba’s new flagship model for code and cowork
Alibaba has unveiled Qwen3.8-Max, described as the most capable model the company has produced to date, with a focus on code and collaborative work with AI (cowork). The model has 2.4T parameters and is available immediately via the API and Qwen Studio.
The team highlights four areas of progress: long-duration autonomous development (ten days or more of self-evolution, from an empty folder to production, with a complete project trace published on GitHub), production-quality deliverables across many roles, autonomous planning over long horizons — up to 500 rounds of chip-design optimization and 365 days of e-commerce strategy cited as examples — and native multimodal intelligence where vision continuously feeds planning, execution, and self-correction.
Alibaba says the weights for Qwen3.8-Max will be released the week after the announcement, along with open weights for a lighter version, Qwen3.8-27B. The model is also available from the announcement on Venice, an anonymous chat platform, and an integration into a tool called “Command Code” was mentioned the same day without additional details.
| Indicator | Value |
|---|---|
| Parameters | 2.4T |
| Input price | $2.0 / million tokens |
| Output price | $6.0 / million tokens |
| Implicit cache | $0.25 / million tokens |
| Weight release | Following week (Qwen3.8-Max + Qwen3.8-27B) |
Genspark opens the source code of GenOffice, its AI office suite
Genspark announced that GenOffice is going open source, presented as the first complete open-source AI office suite for PC and Mac. The tool is free for everyone and ad-free.
The suite covers the classic uses of an office suite — word processing (Docs), spreadsheets (Sheets), presentations (Slides), and PDF — with the usual editing tools. Genspark integrates its Super Agent, capable of searching, analyzing data, then drafting a document or building a presentation, while consuming Genspark credits.
A notable development-method detail: the Alpha version was built by a single engineer in one week, at a cost of $10,000 in tokens.
“One engineer, one week, $10,000 in tokens. That’s what it took to build the GenOffice Alpha.” — @genspark_ai on X
The project is in an openly acknowledged Alpha phase: Genspark invites the community to test the tool, report bugs, and propose feature requests via a dedicated group chat on GenTeam, without writing code. The most helpful contributors receive more than 1,000 Genspark credits as thanks. The announcement took place at the Singapore edition of AGI Playground, where Genspark argues that as models become cheaper, value shifts toward the best experience built on top.
Vercel launches a programmable API for v0
Vercel is opening a new v0 API, available starting today, that gives programmatic access to app-building capabilities previously reserved for the chat interface. In practical terms, a developer can now drive v0 from their own code rather than going through the website.
Four operations are exposed: start a conversation from a text prompt, an existing code repository, or a ZIP archive; generate a development server preview to visualize the result live; send follow-up messages to iterate on the generation; and deploy the result directly on Vercel.
This API turns v0 into a building block that can be embedded in third-party toolchains — agents, automation scripts, internal integrations — rather than just a standalone consumer product. It is a coherent evolution in Vercel’s strategy to make v0 a UI-generation layer that can be used by other agents, including third-party agents that drive v0 without going through its original chat.
For teams already building development agents, this opens the possibility of chaining code generation, preview, and deployment in a fully automated pipeline, without human intervention at each step.
Cursor: more efficient cloud agents and new Google Workspace plugins
Cursor published two separate updates on the same day, one on the performance of its cloud agents, the other on Google Workspace integration.
Cloud agents: 20 to 30% fewer tokens
Cursor announces significant efficiency gains for its cloud agents: 20 to 30% fewer tokens consumed across all tasks, and up to 80% more efficiency on tasks using computer control (computer use). These gains come from better management of MCP servers, skills, and computer use by the agents. The stated goal is to let users delegate more ambitious tasks while staying within their token budget.
Google Workspace plugins integrated into the editor
Cursor can now read, write, and act directly on Google Workspace via plugins accessible from the “Customize” page or the Cursor Marketplace, without leaving the editor.
| Workspace service | Features |
|---|---|
| Google Drive | Search, open, and download files, organize folders |
| Gmail | Search and read mail, draft and send messages |
| Google Calendar | View calendars, create events, find open slots |
| Google Docs | Open, read, write, and edit documents |
| Google Sheets | Read ranges, update cells, create sheets |
| Google Chat | Access spaces and messages, send messages |
The goal is to let a developer pull context from their mail or calendar directly from Cursor, without back-and-forth trips to the browser.
GitHub Agentic Workflows automates cross-repo documentation at Aspire
GitHub highlights a concrete use case for GitHub Agentic Workflows: the Aspire team uses it to automatically turn merged features into documentation pull requests spread across multiple repositories. The system applies scoped permissions and requires a human review by a subject-matter expert (SME review) before merge.
| Metric tracked | Value |
|---|---|
| Documentation PRs merged | 82 / 82 |
| Median merge time | 44.8 hours |
This case study illustrates one of GitHub’s priority areas for agentic workflows: keeping a human in the loop while automating repetitive documentation maintenance tasks.
GitHub publishes its July 2026 Ship Log
GitHub publishes its monthly roundup of July 2026 releases, centered on the theme of choice for developers.
The GitHub Copilot desktop app is now available for all plans, including Copilot Free and GitHub Education; developers without a subscription can also use their own model provider key (bring your own key). Copilot’s model selector is expanding: Kimi K2.7 Code becomes the first open-weights model selectable in the picker, the OpenAI GPT-5.6 family joins Copilot with price cuts of 80% for GPT-5.6 Luna and 20% for GPT-5.6 Terra, and Anthropic’s Opus 5 reaches general availability. GitHub Projects also gains improved filters and a tag system based on multi-select fields.
Sakana Namazu, an LLM API tailored to the Japanese market
Sakana AI has launched the Sakana Namazu API, a service version of its Namazu model that had previously been reserved for the Sakana Chat app. The model is based on Kimi K2.6, Moonshot AI’s open-weights model, refined by Sakana AI for the Japanese language and business use cases.
The API is compatible with the OpenAI format: you only need to change the base URL and API key in existing code. It includes built-in web search and code execution tools, with agentic capabilities to chain multiple steps autonomously.
Sakana AI highlights three use cases: autonomous generation of market research reports, customer support automation, and a creative application controlling about 1,000 animated fish from a simple instruction. On the FairPoliticsQA benchmark, dedicated to Japanese cultural context, the score rises from 34.10% to 56.30%.
MiniMax H3: the weights are now published
MiniMax had announced its open multimodal H3 model on July 31 — already covered in these pages. On August 3, the weights are actually released as open weights, with claimed “Day-0” support across the entire open-source inference stack.
| Inference tool | Day-0 support |
|---|---|
| ComfyUI | Native integration, one model and five workflows from launch |
| vLLM-Omni | OpenAI API-compatible video endpoint |
| SGLang / LMSYS | Customizable local support, execution on NVIDIA and AMD hardware |
| fal | Production infrastructure ready to use |
MiniMax presents this release as crossing a milestone between the announcement and actual availability: the weights are open, and production readiness is too. Third-party partner Magnific also presented a character-based native lip-sync demo, in a single generation.
Runway launches a Max plan with 7 days of unlimited access to Seedance 2.5
Runway announces a new subscription tier, the Max plan, with a launch offer: subscribers get 7 days of unlimited access to the Seedance 2.5 model as soon as it becomes available on the platform.
Seedance 2.5, already mentioned in our previous editions, is not yet available on Runway (“coming soon”): it is the new plan and its early-access offer that are the day’s novelty, not the model itself.
ElevenAgents: ElevenLabs’ voice AI at enterprise scale
ElevenLabs details the scale and capabilities of its voice and chat agent platform, ElevenAgents.
“More than 10 million conversations run on ElevenAgents every week.” — More than ten million conversations are handled each week on ElevenAgents.
Each conversation corresponds to a real interaction handled by a natural-sounding voice or chat agent: refund request, appointment booking, question about benefits, flight change. The platform offers scenario simulation before production, multichannel deployment, and monitoring via the Spotlight tool, which watches conversations in production and suggests proactive improvement recommendations. On the enterprise side, ElevenLabs emphasizes configurable safeguards and regional data-residency options, for customers ranging from startups to governments and Fortune 500 companies.
Google and PSG: Gemini becomes the club’s official AI assistant
Paris Saint-Germain and Google have announced a Premium partnership running through 2029. Under this agreement, Google Gemini becomes the club’s official AI assistant, while Google Pixel becomes PSG’s official smartphone.
This is a sponsorship and distribution deal, with no technical details at this stage about any associated features or product integrations. The move fits into Google’s strategy to broaden Gemini’s mainstream visibility beyond its circle of tech-savvy users by associating it with one of the world’s most-followed football clubs.
Briefs
- Vercel and v0 now accept sign-in with a ChatGPT account — a “Continue with ChatGPT” button has been added to the login pages, in addition to the Vercel plugin now available in ChatGPT. 🔗 Source
- Replit announces the Designathon — a one-week design contest open to all creators, with a live launch on August 4 at 9 a.m. Pacific time and prizes up for grabs. 🔗 Source
- Together AI compares the cost per task of Kimi K3 Max and Fable 5 xhigh — on the DeepSWE benchmark, Kimi K3 Max approaches the performance of Fable 5 xhigh for about one third of the per-run cost, or 2.8 times more tasks solved per dollar spent. 🔗 Source
- Migration from GitLab to GitHub Enterprise Cloud reaches general availability — the
gh gl2ghCLI extension now makes it possible to self-serve migrate repositories from gitlab.com or GitLab Self-Managed to GitHub Enterprise Cloud. 🔗 Source - Luma Ray 3.2 available on third-party platform Fuser — the video model is integrated into this cropping tool, with the promise of preserving the scene’s action when changing aspect ratio. 🔗 Source
- NVIDIA NeMo Gym integrates Legal Agent Bench — this legal-agent benchmark, developed with Harvey, joins the open NeMo Gym library to advance open-weights models on complex legal use cases. 🔗 Source
- NVIDIA demo of a parkour agent learned in 30 seconds — a simulation reportedly acquired expert behavior from a very short reference video excerpt, with no further technical details provided. 🔗 Source
- Kimi Slides tutorial for generating presentations — Moonshot AI publishes the first episode of an educational series on Kimi Slides, presentation generation powered by Kimi K3, with charts and SmartArts. 🔗 Source
What it means
The race for open weights is moving forward on several fronts at once. Qwen3.8-Max, despite its 2.4T parameters, will be open as soon as next week; MiniMax H3 is moving from announcement to real availability with Day-0 support across four different inference stacks; and Together AI’s cost/performance analysis of Kimi K3 Max illustrates why these open models are gaining ground: at similar quality, they cost a fraction of the price of equivalent closed models. The balance of power between open and closed models is increasingly being decided on this economic ground rather than on benchmark scores alone.
The second axis is the platformization of development tools into programmable building blocks. Vercel’s v0 API, the efficiency gains from Cursor’s cloud agents and Google Workspace integration, and the use case documented by GitHub at Aspire all point to the same trend: vendors are no longer selling just a chat interface, but an automation chain that can be driven by other agents, with guardrails — restricted permissions, human review — to make it usable in production.
On the consumer side, two announcements operate in a different register: Genspark’s open-sourcing of GenOffice aims to democratize a full AI office suite by leaning on free access and community effects, while the partnership between Google and PSG is more about brand visibility than technical innovation — Gemini gains mainstream exposure without any new feature being announced.
Finally, the figures cited by ElevenLabs — more than ten million weekly conversations on ElevenAgents — and Runway’s launch offer around Seedance 2.5 confirm that generative AI for production, both voice and video, is now being deployed at enterprise scale, with dedicated oversight tools rather than simple technology demos.
Sources
- Qwen3.8-Max — announcement on X
- GenOffice — announcement on X
- API v0 — announcement on X
- Cursor — cloud agents on X
- Cursor — Google Workspace plugins
- GitHub Agentic Workflows at Aspire
- GitHub Ship Log July 2026
- Sakana Namazu — official announcement
- MiniMax H3 — published weights
- Runway — Max plan
- ElevenAgents — ElevenLabs on X
- Google and PSG — announcement on X
- Vercel and v0 — ChatGPT connection
- Replit Designathon
- Together AI — Kimi K3 Max vs Fable 5 xhigh
- Migration from GitLab to GitHub Enterprise Cloud
- Luma Ray 3.2 on Fuser
- NVIDIA NeMo Gym — Legal Agent Bench
- NVIDIA — parkour demonstration
- Kimi Slides — tutorial