~8m21:47DeepSeek V4 Flash Explained: Why It's Worth Switching To Right Now (+14 AI Updates)
Aug 9, 2026
Read: ~8m · You save: 14 min
DeepSeek V4 Flash Explained: Why It's Worth Switching To Right Now (+14 AI Updates)
Meta's Muse Code, Alibaba's Qwen 3.8, GPT 5.6 price cuts, a free AI office suite, and DeepSeek V4 Flash. Plus a hands-on Muse Spark 1.2 tutorial.
August 9, 2026
Major AI developers and technology companies released several products and updates this week. The releases span office software, large language models, robotics, video generation, and coding agents.
Productivity Software
GenSpark GenOffice
GenSpark released GenOffice, which it describes as the world's first open-source AI office suite for PC and Mac. Each tool includes a built-in AI assistant. The suite competes with Google Workspace and Microsoft Office. Microsoft's most expensive Office plan costs 20,000 rupees, or $210.
- Documents: A prompt such as "Create a professional resume for a Stanford CS master's graduate" produces a complete resume on the page. Users can then edit it with standard formatting tools.
- Spreadsheets: Users can open a large data set and ask a question. GenOffice analyzes the data, builds a summary table, and supports formulas that are standard in Excel. It can also generate charts from the data.
- Presentations: A prompt such as "Polish this slide while keeping the same colors and theme" restructures the layout and improves readability.
- PDFs: A prompt such as "Highlight what is worth reading" marks key sections and produces a summary of the main findings.
The suite itself is free. Users pay only for the AI models they use.
Chinese AI Models
Alibaba Qwen 3.8
Alibaba released Qwen 3.8, which has 2.4 trillion parameters. Alibaba states that this matches the parameter count of leading models from OpenAI and Anthropic. The company reports that Qwen 3.8 outperforms top models on reasoning benchmarks, research tasks, and computer use tests.
Alibaba describes the model as an "always-on workmate." Documented use cases include:
- Designing computer chips
- Verifying company finances
- Generating contracts automatically
- Drafting design proposals
Qwen 3.8 is built for long-running tasks that continue for hours without new instructions. In one example, Qwen 3.8 Max wrote code on its own for 16 days. It tested and improved the same project without constant human input. It also added major features, including a goal system, a full desktop application, and an interactive galaxy map.
DeepSeek V4 Flash
DeepSeek released the beta version of V4 Flash. The model targets AI agents that write code, use tools, and complete long tasks.
- Context window: 1 million tokens, which allows a single conversation to hold large volumes of information.
- Pricing: 0.28 per million output tokens.
DeepSeek V4 Pro is the more powerful version in the same family.
Kimi K3
Moonshot's Kimi K3 gained rapid popularity last month. Demand exceeded available hardware, and the company paused new sign-ups.
OpenAI
GPT-5.6 Price Changes and Model Versions
GPT-5.6 has three versions:
| Version | Role |
|---|---|
| Soul | Most powerful |
| Terra | Balance of intelligence and cost |
| Luna | Fastest |
OpenAI cut API costs for Luna by 80% and for Terra by 20%. According to OpenAI's comparison graph, Luna costs less than US models and less than Chinese models, including GLM 5.2 and MiniMax.
OpenAI states that GPT-5.6 Soul helped optimize the systems that run it, which reduced its own operating cost.
Two additional changes apply to the model family:
- Fast mode: GPT-5.6 Soul offers fast mode, which runs 2.5 times faster than standard processing. Intelligence does not change.
- Free access: GPT-5.6 Luna offers unlimited text chats on the free plan.
ChatGPT Chrome Extension Updates
OpenAI released three upgrades to the ChatGPT Chrome extension:
- Video analysis: Users open a YouTube video and request a summary. ChatGPT produces it within seconds. Previously, users had to paste the transcript manually.
- Tab references: Users type "@" to mention any open tab. ChatGPT can compare a video with open articles, documents, or guides and produce one combined answer.
- Text editing in web pages and Google Docs: Users highlight text on a web page or in a Google Doc and discuss that section with ChatGPT. It can edit the section directly in the Google Doc.
Gemini Spark
Gemini Spark operates websites on the user's behalf. A user can give one instruction, such as: "Find the cheapest non-stop flight, log into my account, fill in my details, and stop when it is ready to book."
After the user grants permission, Gemini Spark opens the travel site, checks airports, enters dates, filters results, and selects the lowest-cost option. The user can monitor each step. At the final booking step, Gemini Spark stops and returns control to the user.
Gemini Robotics 2
Google DeepMind released Gemini Robotics 2, a physical AI model that controls an entire humanoid robot, from the body to the arms and attached tools. The model enables three capabilities:
- Whole-body control: Robots must coordinate arms, legs, and balance at the same time. Packing a box requires this coordination.
- Dexterous hands: Hand movement requires coordinating 22 joints. Holding a plastic bag requires control of all of them together.
- Robot-to-robot coordination: One robot can give instructions to another. In a demonstration, a command to a robot named Duo was: "Get all tools in the bin, close the kit, and put the kit back into the bin." Duo responded, "On it." The robots reason through tasks and coordinate to complete workflows.
Meta
Muse Code and Muse Spark 1.2
Meta released Muse Code, a coding agent that competes with Codex and Claude Code. Muse Code runs on Muse Spark 1.2, a model built for real coding workflows and longer tasks. Muse Spark 1.2 has a 1 million token context window.
Muse Code supports the following functions:
- Planning software tasks
- Writing code
- Testing code to confirm that it works
Its main feature is memory. Background agents stay active throughout a session and build context over time. They do not start from zero on each task. For large jobs, Muse Code divides work among sub-agents that run in parallel. In one test, it built six features simultaneously without conflicts. If a task crashes, Muse Code resumes from the point of failure, so no work is lost and no prompts need to be repeated.
According to Meta, Muse Code's performance is close to GPT-5.6, Gemini 3.6 Flash, and Opus 5. It is not the leading model at present. Muse Code is in beta. Meta demonstrates it for game development, web development, and multimodal visual coding.
Muse Code Setup
- Open the Muse Code page and copy the installation command.
- Open a terminal, paste the command, and allow installation to finish.
- Type
museto launch the application. - On first launch, enter an API key or log in through a browser. The browser login opens Meta's approval page. Approve Muse Code and return to the terminal.
After launch, type a slash to list the available commands. Two commands are useful:
/effort: Sets the level of reasoning. Options are minimal, low, medium, high, extra high, and ultra./model: Switches between Muse Spark 1.1, Muse Spark 1.2, and Muse Spark 1.2 contributor.
Pricing and Data Use
Pricing ranges are as follows:
| Model | Price range |
|---|---|
| Muse Spark 1.2 | 4.25 |
| Muse Spark 1.2 contributor | 0.2 |
Meta states that content submitted to the contributor model may be used for product improvement. The contributor model suits personal projects. Users working with private company code or sensitive material should select the standard model.
Test Projects
The author tested Muse Spark 1.2 on three projects:
- Spider-Man 3D character in Blender: The author provided a Spider-Man reference image and asked Muse Code to recreate it as a detailed 3D character using Blender MCP. Muse Code produced a 3D model that can be rotated, edited, animated, and exported.
- Scroll-based landing page: The author provided a space fly-through video, existing images, and assets. The request was a cinematic landing page in which the video moves with the user's scrolling. Muse Code built the layout, typography, sections, and interactions around the video.
- Monza Ultimatum browser game: The author requested a Formula 1-style browser game. The specification covered the environment, racing track, car physics, controls, and gameplay. The output included a starting light sequence, a drivable 3D car, live speed readout, lap timing, position tracking, and a mini-map.
Video Generation: ByteDance Seedance 2.5
ByteDance released Seedance 2.5, a video model that generates animated and realistic video.
- Audio and video: The model generates 30 seconds of video and audio together, which keeps voices and movements synchronized.
- Character consistency: A character can move across shots while keeping the same face, voice, and lighting.
- Scene handling: The model manages dialogue and changes in camera angle without breaking the scene.
- Physics: The model renders real-world physical behavior, such as rain and moving engineering parts.
- References: Each prompt accepts up to 30 images, 10 video clips, and 10 audio clips.
Design and Workspaces
Replit Design
Replit released Replit Design, a tool for creating designs without design experience. Users can start from templates or from a new direction.
- New direction: The tool offers options such as a ticket checkout flow, an exhibition gallery, or a mobile app. In the author's test, an exhibition gallery produced sections for a press kit, contact page, and event invitations.
- Templates: Hundreds of ready-made templates are available.
- Brand consistency: The "apply design system" function applies one visual style across all screens.
- Output: Users can create social media posts and websites in one step. Replit states that the tool uses leading US and Chinese models, including models from OpenAI and Kimi.
Perplexity Projects
Perplexity launched Projects, a workspace for ongoing work. Standard AI chats start from zero each time. Projects keeps files, connected apps, and specialized skills in one workspace.
- Connected apps: Slack, Gmail, and Google Drive connect to a project.
- Structured queries: A question such as "What are the outstanding reps and warranties in the purchase agreement?" produces a structured summary and next steps based on the project's files.
- Team access: A team member can tag Perplexity in Slack and request a change. The resulting task appears on the project dashboard.
- Document work: Projects can draft documents, build presentations, and work across spreadsheets while keeping the original context.
Platforms
LinkedIn "Seems Like AI Slop" Button
LinkedIn introduced a "seems like AI slop" button. Users can report a post that appears fake, repetitive, or copied from a chatbot.
A study of more than 1 million social media posts found that 41% of long-form LinkedIn posts were fully AI-generated. When enough users flag a post, it is removed from the feed. LinkedIn may also inform the author privately that the writing appears inauthentic.
Buzz
Jack Dorsey, former CEO of Twitter, is building Buzz, a group chat platform that competes with Slack and GitHub. Buzz allows humans and AI agents to work in the same team chat. Users can share updates, assign work, and track project progress without switching apps. Buzz is open source.
A demonstration shows three agents, named Bumble, Fizz, and Honey, completing a task in sequence:
- Bumble creates the prototype and asks Fizz to handle deployment.
- Fizz deploys the prototype and asks Honey to review it.
- Honey completes the review and asks a human for final approval.
The human makes the final decision.
Agents and Research
Hermes "Herald" Release
Nous Research released the Herald release of its open-source Hermes agent. The update includes five features:
- Voice interaction: Hermes speaks while a response loads and stops when the user interrupts. It can respond to a custom wake phrase, such as "Hey Hermes," and end a conversation on the command "Stop."
- Source verification: Hermes links each claim to a source, compares quotations with the source text, and runs a fact-check mode that identifies claims that do not hold up.
- Agent collaboration: Hermes agents can communicate with each other and transfer work.
- Response speed: Responses arrive in 0.9 seconds, compared with 4.3 seconds before the update. On desktop, responses are eight times faster.
- Mid-task correction: Users can correct Hermes while it works, and the agent changes course without a restart.
OpenAI Astra and Mathematical Proofs
OpenAI's upcoming model, Astra, solved 10 complex mathematical problems that had remained unsolved for decades. The problems involved cryptography and combinatorics. The final results reportedly required about $2,000 in model usage. OpenAI released the proofs so that mathematicians can inspect and verify them.
SpaceX Starmind
SpaceX is developing Starmind, a project to run AI computers in orbit using large satellites. Each satellite would:
- Measure 70 meters across
- Generate up to 250 kW of power
- Use large solar panels to collect near-constant sunlight
Results would reach Earth through Starlink's high-speed laser network. SpaceX plans to use Nvidia GPUs and CPUs for compute. Early deployments are planned for as soon as 2027.
Introduction: Meta, China, and the AI Race
Meta launches Muse Code, China's AI labs advance, and DeepSeek V4 Flash is pitched as the cheapest major model.
- Meta launched Muse Code, a rival to Claude Code, Antigravity, and Codex, able to make software, 3D models, and games
- Alibaba launched Qwen 3.8, described as trained on 2.4 trillion parameters
- Kimi K3 reportedly ran out of hardware and paused sign-ups last month
- DeepSeek launched V4 Flash, described as the cheapest major AI model running today
GenSpark GenOffice
GenSpark released GenOffice, a free open-source AI office suite for PC and Mac with built-in AI assistants in each tool.
- GenOffice is described as the world's first open-source AI office for PC and Mac
- Documents: AI writes full resumes from prompts, such as a resume for a Stanford CS master's graduate
- Spreadsheets: AI analyzes large data sets, builds summary tables, supports Excel-like formulas, and creates charts
- Presentations: AI restructures slides while keeping colors and theme; PDFs can be highlighted for key sections
- The tool is free; users pay only for the AI they use, versus Microsoft's premium plan at 20,000 rupees ($210)
Alibaba Qwen 3.8
Alibaba's Qwen 3.8 is pitched as an always-on AI workmate that works on long tasks for hours, including a 16-day autonomous coding run.
- Qwen 3.8 is described as trained on 2.4 trillion parameters, comparable to top OpenAI and Anthropic models
- Claimed to beat top models on reasoning benchmarks, research, and computer use
- Use cases include designing computer chips, verifying company finances, generating contracts, and drafting design proposals
- Qwen 3.8 Max reportedly kept coding on its own for 16 days, self-testing and improving the same project
- Added features in the example include a goal system, a full desktop app, and an interactive galaxy map
OpenAI Price Cuts and GPT 5.6
OpenAI cut GPT 5.6 prices, added a fast mode, and made Luna unlimited for free users in response to China's competition.
- GPT 5.6 has three versions: Soul (most powerful), Terra (balance of intelligence and cost), and Luna (fastest)
- Luna's API cost was cut by 80% and Terra's by 20%
- Luna is claimed to be cheaper than some US and Chinese models such as GLM 5.2 and MiniMax
- GPT 5.6 Soul reportedly helped optimize the systems running it, cutting its own cost
- Fast mode for GPT 5.6 Soul offers 2.5 times faster speeds with no change in intelligence
- GPT 5.6 Luna offers unlimited text chats for everyone, including free plans
ChatGPT Chrome Extension and Gemini Spark
ChatGPT's Chrome extension gained fast video analysis, tab mentions, and in-page text highlighting and editing. Google's Gemini Spark can operate websites for users.
- Video analysis: open a YouTube video and ask for a summary in seconds
- Typing @ mentions any open tab to combine content from multiple sources into one answer
- Highlighting text on a web page or Google Doc lets ChatGPT discuss or edit that section directly
- Gemini Spark can use websites, such as finding the cheapest non-stop flight, filling details, and stopping before booking
- Users can watch each step, and Gemini hands control back at the final booking step
Meta Muse Code
Meta's Muse Code is a coding agent powered by Muse Spark 1.2, featuring persistent background agents, parallel sub-agents, and crash recovery.
- Muse Code is powered by Muse Spark 1.2 and competes with Codex and Claude Code
- It plans, writes, and tests software tasks
- Background agents stay active across the session to build context over time
- Sub-agents work in parallel; a demo game built six features simultaneously with no collisions
- If it crashes mid-task, it resumes exactly where it left off
- Performance is described as close to GPT 5.6, Gemini 3.6 Flash, and Opus 5, but not the leading model yet
Google DeepMind Gemini Robotics 2
Google DeepMind's Gemini Robotics 2 is a physical AI that controls a humanoid's whole body, dexterous hands, and multi-robot coordination.
- Gemini Robotics 2 controls a humanoid from its body to arms and attachments
- Whole-body control coordinates arms, legs, and balance together
- Dexterous hands coordinate 22 joints
- Robots can give instructions to one another, as demonstrated with a tool-clearing task
ByteDance Seedance 2.5
ByteDance's Seedance 2.5 generates 30 seconds of synchronized video and audio, supporting animated and realistic styles and multi-shot consistency.
- Seedance 2.5 creates both animated and realistic videos
- Characters keep the same face, voice, and lighting across shots
- Generates 30 seconds of video and audio together to keep voices and movements in sync
- Supports up to 30 image, 10 video clip, and 10 audio clip references per prompt
Replit Design
Replit Design generates design sections from templates or new directions and applies a design system across screens.
- Options include a ticket checkout flow, exhibition gallery, or mobile app
- Creates sections such as press kit, contact, and event invites with one click
- Hundreds of ready-made templates are available
- Apply design system keeps visual style consistent across screens
- Powered by top US and Chinese models, including OpenAI and Kimi
Perplexity Projects
Perplexity Projects provides a persistent workspace with files, connected apps, and skills, integrated into team tools like Slack.
- Projects keeps files, connected apps (Slack, Gmail, Drive), and skills in one ongoing workspace
- Example query: outstanding reps and warranties in a purchase agreement, returned as a structured summary with next steps
- Teammates can tag Perplexity in Slack, and tasks appear on the project dashboard
- Can draft documents, build presentations, and work across spreadsheets while keeping context
LinkedIn AI Slop Button
LinkedIn added a 'seems like AI slop' report button; a study found 41% of long-form posts were fully AI generated.
- Users can report posts that feel fake, repetitive, or copied from ChatGPT with one tap
- A study of over 1 million social media posts found 41% of long-form LinkedIn posts were fully AI generated
- Flagged posts that get enough reports disappear from the feed; creators may be privately told the writing seems inauthentic
Jack Dorsey's Buzz
Jack Dorsey's Buzz is an open-source team chat platform where humans and AI agents collaborate, challenging Slack and GitHub.
- Buzz allows humans and AI agents to work together in the same team chat
- Demo agents Bumble, Fizz, and Honey pass tasks: prototype, deploy, review, then a human final review
- Users can share updates, assign work, and track project progress without switching apps
- The entire platform is open source
Nous Research Hermes Herald
Nous Research's Hermes Herald release adds natural voice conversation, source-linked facts, faster responses, and mid-task correction.
- Voice: speaks while loading, stops on interruption, wakes on 'Hey Hermes', and ends on 'Stop'
- Claims link to verifiable sources, quotes are matched against page text, and a fact-check mode flags unsupported claims
- Hermes agents can talk to each other and hand over work
- Responses arrive in 0.9 seconds versus 4.3 seconds before; desktop answers are eight times faster
- Users can correct answers mid-task and the agent redirects
OpenAI Astra Math Results
OpenAI's upcoming model Astra reportedly solved 10 long-standing complex math problems for roughly $2,000 in model usage, with proofs released for verification.
- Solved 10 complex math problems unsolved for decades, including cryptography and combinatorics
- Final results reportedly used about $2,000 worth of model usage
- OpenAI released proofs so mathematicians could inspect and verify them
SpaceX Starmind Orbital AI
SpaceX's Starmind project plans giant orbital satellites running AI computers, powered by solar panels and linked to Earth via Starlink lasers.
- Each AI satellite could stretch 70 m across and generate up to 250 kW of power
- Uses massive solar panels for near-constant sunlight
- Results travel to Earth through Starlink's high-speed laser network
- Built with Nvidia GPUs and CPUs; early deployments hoped for as soon as 2027
DeepSeek V4 Flash
DeepSeek V4 Flash beta targets AI agents with a 1 million token context window and very low pricing.
- Beta version built for agents that write code, use tools, and handle long tasks
- 1 million token context window
- Costs 14 cents per million input tokens and 28 cents per million output tokens
- DeepSeek V4 Pro is the more powerful version
Muse Spark 1.2 Tutorial
Tutorial on Muse Code setup, model options, and pricing, plus three tests: a Blender Spider-Man model, a scroll-based landing page, and a Formula 1-style game.
- Install by copying the command from the Muse Code page, running it in terminal, and typing 'Muse'
- Login via browser approval or API key
- Slash commands: /effort (minimal, low, medium, high, extra high, ultra) and /model (Muse Spark 1.1, 1.2, 1.2 contributor)
- Standard Muse Spark 1.2 costs 4.25; contributor version costs 0.2
- Contributor model content may be used for product improvement, so avoid it for private or sensitive code
- Project 1: Spider-Man 3D character in Blender via Blender MCP from a reference image
- Project 2: Staying Ahead website with a video-driven smooth scroll hero animation
- Project 3: Monza Ultimatum browser game with starting lights, 3D car, speed, lap timing, position, and mini-map