Blog

Summary: Perplexity Spaces — previously called Collections — are folders for your AI threads with a superpower: you can attach custom instructions and files that every answer inside that Space follows. Instead of re-explaining what you want in every chat, you set the rules once. This guide shows you how to set up a Space, write instructions that actually work, and includes copy-paste templates for research, coding, studying, and market research.

What Is a Perplexity Space?

A Space is a container for related threads. Inside a Space you can:

• Group threads on one topic so your research stays organised
• Add custom instructions — persistent directions the AI follows for every answer in that Space
• Attach files the AI can reference when answering
• Share the Space with others so they can ask questions inside it with the same context

Think of it this way: a normal Perplexity thread starts from zero every time. A Space starts from your instructions every time.

Note: Perplexity renamed Collections to Spaces, so older guides that say “Collections” mean the same feature. Menu labels can shift between updates — if a button has moved, look for “Spaces” in the sidebar or Library.

How to Set Up a Space (Step by Step)

Step 1: Create the Space

Open Perplexity, go to your Library or the Spaces section in the sidebar, and create a new Space. Give it a clear, specific name — “AI Video Tools Research” beats “Research” — because the name reminds you (and collaborators) what belongs inside it.

Step 2: Write a short description

Add one or two lines describing what the Space is for. This is mostly for you and your collaborators, but a clear description keeps the Space focused over time.

Step 3: Add custom instructions

This is the important step. Custom instructions are a short block of text that tells the AI how to behave in this Space: what role to take, what format to use, what to prioritise, and what to avoid. Keep them specific and short — a tight paragraph beats a long essay the model will half-follow.

Step 4: Attach reference files (optional)

If you have documents the AI should use — a brief, a spec sheet, a set of notes — attach them to the Space. Answers inside the Space can then draw on those files instead of only on the web.

Step 5: Start threads inside the Space

New threads created inside the Space inherit its instructions automatically. Threads you started elsewhere do not pick up the Space’s instructions if you move them later — when in doubt, start fresh inside the Space.

How to Write Custom Instructions That Actually Work

Good instructions share four traits:

1. A role: “You are a research assistant for a tech blog” tells the model what job it is doing.
2. An output format: “Answer with a short summary first, then bullet points” prevents walls of text.
3. A quality rule: “Cite sources for factual claims” or “Flag anything you are unsure about instead of guessing.”
4. A boundary: “Do not recommend paid tools unless the free option genuinely cannot do the job.”

Bad instructions are vague (“be helpful”), contradictory (“be brief but cover everything”), or so long that the model ignores half of them. Aim for 4–8 lines.

Copy-Paste Instruction Templates

Template 1: Research Assistant

You are a research assistant for a technology blog. For every question, start with a 2–3 sentence summary, then give the key facts as bullet points with sources. Separate confirmed facts from rumours or leaks, and label them clearly. If sources disagree, say so and explain both sides. Keep the tone plain and practical. Do not hype products.

Template 2: Coding Helper

You are a coding assistant. Give working code first, then a short explanation of how it works. Prefer simple, readable solutions over clever ones. When a question has multiple valid approaches, show the simplest one and mention the alternative in one line. If code could fail (missing dependency, version issue), say so explicitly.

Template 3: Study Notes

You are a study tutor. Explain concepts in plain language with one concrete example each. After each explanation, add a one-line “remember this” summary. Ask me one quick check question at the end of each answer so I can test myself. Do not dump textbook definitions without examples.

Template 4: Market / Product Research

You are a market research assistant. Compare options in a table with price, key features, and honest limitations. Always include the free or cheapest viable option. Flag when a claim comes from the vendor’s own marketing rather than independent testing. End with a one-line bottom-line recommendation and who it is NOT for.

Honest Limits: What Spaces Can’t Do

• Instructions are guidance, not guarantees. The model can still drift, especially on long threads. If an answer ignores your instructions, restate the key rule in that thread.
• A Space doesn’t change the underlying model. Custom instructions shape style and format; they don’t give the AI new knowledge or make it browse differently than usual.
• Old threads don’t inherit new instructions automatically. If you rewrite a Space’s instructions, start a new thread to feel the full effect.
• Features and plan limits change. Some Space features (like collaboration or file limits) differ between free and paid plans, and Perplexity updates them over time. If a button or limit looks different from this guide, check Perplexity’s current plan page rather than assuming the guide is wrong.

A Practical Workflow: Spaces for a Research Project

1. Create one Space per project, not one Space per question.
2. Write the instructions once, using a template above as a starting point.
3. Attach the 2–3 documents that matter most — not your entire hard drive.
4. Run your first 3–4 threads inside the Space and check whether answers actually follow the instructions. If not, tighten the instructions.
5. When the project ends, leave the Space as-is. It becomes a searchable archive of that project’s research.

Bottom Line

Perplexity Spaces are worth setting up the moment you ask more than two or three questions on the same topic. Ten minutes spent writing good custom instructions saves you from re-explaining yourself in every thread — and the copy-paste templates above give you a starting point you can adapt in minutes.

Frequently Asked Questions

Are Spaces and Collections the same thing?

Yes. Perplexity renamed Collections to Spaces. Older guides that mention Collections refer to the same feature.

Do custom instructions apply to old threads?

Threads started inside the Space follow its instructions. Threads started outside it don’t automatically pick them up — start a new thread inside the Space to be safe.

Can I share a Space with someone else?

Yes, Spaces can be shared so collaborators can ask questions with the same instructions and files. Exactly who can join and what they can edit depends on your plan.

Do Spaces make answers more accurate?

They make answers more consistent and better formatted, and attached files ground answers in your own documents. They don’t make the model infallible — always verify important facts, especially prices, dates, and specifications.

Research used to mean juggling a dozen browser tabs, skimming dense reports, and hoping you didn’t miss the one paragraph that mattered. AI search tools changed that, and Perplexity has become one of the most useful ones for anyone who needs answers they can actually trace back to sources. But most people use it like a basic search box and miss the features that make it genuinely powerful. Here is a practical workflow for doing real research with Perplexity, from first question to finished notes.

Ask the right questions first

Perplexity answers whatever you type, so the quality of your research starts with the quality of your prompt. Vague questions produce vague, generic summaries. Instead of “tell me about renewable energy,” ask “What were the fastest-growing renewable energy sources in India in 2025, and what policy changes drove that growth?” Specific questions give the model concrete terms to search for and anchor its answer to real data points.

When you are exploring a topic you don’t fully understand yet, start broad on purpose and let the answer teach you the vocabulary. Jot down the key terms, statistics, and names that appear, then run a second, sharper query using them. This two-pass approach — an exploratory question followed by a precise one — consistently produces deeper, more accurate results than trying to nail it in one attempt.

Build follow-up threads instead of starting over

One of Perplexity’s best habits is the follow-up: after each answer, suggested follow-up questions appear beneath it, and you can also type your own. Treat the first answer as the opening of a conversation, not the final word. Drill down with follow-ups like “What do critics of this approach say?” or “Can you break down that statistic by year?” Each follow-up stays in the same thread, so the context carries over and you avoid repeating background information.

Threads also keep your research organized automatically. A single thread becomes the full arc of one investigation — initial question, refinements, counterarguments, and final synthesis — which makes it much easier to revisit or summarize later than a scattered history of unrelated queries.

Use the sources panel like a researcher, not a skimmer

Every Perplexity answer includes a sources panel with numbered citations linking to the web pages, papers, and reports it drew on. Don’t just read the summary — click through. Open the top three to five sources and check what type of content they are: official reports, academic papers, news outlets, or blogs and forums. A claim backed by an official source carries far more weight than one built on forum threads.

Get in the habit of verifying the single most important claim in any answer before you rely on it. Open the cited source, find the passage, and confirm the numbers and context match. AI tools can occasionally misread a table or conflate two similar statistics, and five seconds of verification is what separates research from guesswork.

Pick the right Focus mode for the job

Perplexity’s Focus modes let you tell it where to search, and choosing well is one of the highest-leverage things you can do. The available modes are typically:

• Web — the default; broad results from across the internet. Best for current events and general topics.
• Academic — searches scholarly papers and journals. Use it for anything technical, scientific, or where you need peer-reviewed backing.
• Social — pulls from discussions and posts. Useful for gauging real-world opinions and experiences.
• Writing — optimized for generating text without searching, for drafting and brainstorming.
• Math — routes through a model tuned for computation and problem-solving.

Match the mode to the question: a market-size query deserves Web, a question about a machine-learning technique deserves Academic. You can also attach files — PDFs, documents, data sheets — and ask questions directly about their contents, which turns Perplexity into a personal reading assistant for long reports.

Verify citations before you quote them

Cited answers are only as trustworthy as the verification behind them, so build a simple verification routine into every research session. First, check the publication date of each key source — AI search sometimes surfaces outdated pages alongside fresh ones. Second, prefer primary sources over articles about articles; if Perplexity cites a news piece about a study, find the study itself. Third, look for corroboration: if two independent sources agree on a figure, you can use it with confidence.

Watch for the classic failure mode of AI research: an answer that sounds confident but cites sources that don’t fully support it. If a citation link leads to a page that doesn’t mention the claimed fact, flag it and ask a follow-up specifically about that claim. Perplexity will often correct or qualify itself when pressed.

Save and organize with Collections

Research generates more threads than you can keep in your head, and that’s what Collections are for. Create a collection per project — a blog post, a purchase decision, a report — and save the useful threads into it. You can also set a custom AI prompt per collection to shape how answers are framed, for example “explain everything at a beginner level” for a learning project or “always include key statistics” for a market report.

Within a collection, Perplexity can search across your saved threads, so a fact you found last week is retrievable instead of lost. For ongoing research, this beats a folder of bookmarks because the content itself is searchable, not just the titles.

A practical workflow: from question to finished brief

Here is how it all fits together in a real session. Suppose you need to decide which AI video generator to recommend for your team. Step one: ask a broad exploratory question in Web focus mode. Step two: note the top three tools mentioned and run a follow-up comparing their pricing and features. Step three: open the sources panel and verify the pricing pages directly, checking the dates. Step four: switch to Academic or Social focus for depth or real user opinions. Step five: save the thread to a “Video tools” collection with a custom prompt asking for comparison tables. By the end, you have a cited, verified, organized brief — in under half an hour, with every claim traceable to a source.

The deeper lesson is that Perplexity rewards researchers, not just questioners. The tool does the fetching and summarizing, but you still do the judging: asking sharply, checking sources, and organizing what you find. Use the workflow above and AI search stops being a shortcut — it becomes a genuine research upgrade.

Related Articles

• Best Free AI Video Generators in 2026: Text-to-Video Tools Compared
• DeepSeek V4.1 Flash 1017: Faster, Leaner and Cheaper — A Practical Guide to the Open-Weight Update
• How to Run FLUX 3 Image Klein: The Open-Weight Image Model That Fits on an 8GB GPU

Turning a sentence into a short video clip with AI used to require a paid subscription. By October 2026, several text-to-video tools give away a meaningful daily or weekly allowance for free. The catch is that every free tier hides real limits: credit caps, watermarks, slower queues and resolution ceilings. Here is an honest comparison of the most useful free AI video generators right now.

The state of free AI video in late 2026

Several well-known tools moved video generation behind paid plans in 2026, and some once-popular apps were discontinued entirely. That makes the remaining genuinely free options worth knowing, because free-plan terms change often and a tool that is generous today can tighten its limits next month.

Most free tiers share a few common restrictions. Clips are usually 5 to 10 seconds long and capped at 720p resolution. Free output is often watermarked, licensed for personal rather than commercial use, and rendered in a slower queue than paid jobs.

Google Flow (Veo 3.1): the best quality free

Google’s Flow platform gives non-subscribers a daily allowance of free credits, reportedly around 50 per day, which is enough for a handful of short clips. The underlying model, Veo 3.1, is widely regarded as the most realistic text-to-video model available, and it can generate synchronized audio, such as dialogue, ambience and sound effects, inside the same clip.

That makes Flow the strongest free option if visual quality is your priority. The trade-offs are real, though: the daily allowance covers only a few generations a day, and availability has been subject to regional limits. Free terms on Google products also shift frequently, so treat the allowance as a snapshot rather than a promise.

Kling: the most generous daily allowance

Kling issues free users a credit allowance on daily login, commonly cited at around 66 credits a day, enough for roughly five or six short clips at standard settings. The model is praised for natural human motion, which is where many AI video generators still look artificial.

The catches are a visible watermark on free exports, a license restricting free output to personal use, and a slower queue for free renders. The newest Kling models also cost more credits per generation, so the daily allowance goes further on standard settings.

PixVerse: the easiest first experiment

PixVerse hands out a small daily credit balance, often reported at around 30 credits a day, with no credit card required. The interface is deliberately simple, which makes it a friendly place for a first text-to-video experiment when you do not want to learn a complex editor.

It is best seen as a low-friction playground rather than a workhorse. The daily balance produces only a few short clips, and for anything longer you will outgrow it quickly. Still, for testing prompt ideas before committing paid credits elsewhere, it is genuinely handy.

CapCut: best for finishing and publishing videos

CapCut is a different kind of tool. It is primarily a free video editor rather than a pure text-to-video generator, but its desktop tier covers full editing at 1080p with exports that carry no watermark on the core workflow. It also bundles AI features such as auto-captions and template effects.

The honest way to use it is as the second half of your workflow: generate raw clips on Kling or Flow, then cut them together, add captions and sound, and export in CapCut. That combination gives you a finished social-ready video without spending anything.

Runway: the pro tool with a one-time trial

Runway’s free plan offers a one-time credit grant, widely reported at 125 credits, that does not renew. It exists so you can test a professional-grade model with advanced camera controls and editing features, not so you can rely on it day to day.

If your goal is to learn the tool before buying, or to produce a few high-quality clips for a one-off project, the trial is useful. Just go in with eyes open: once the credits are gone, they are gone.

What free tiers will not give you

Before you build a workflow around free AI video, be realistic about the boundaries that paid tiers exist to cross:

• Watermark-free output: most free plans brand their exports. Check each platform’s terms before using clips in anything public or commercial.
• Long clips and high resolution: free tiers typically top out at 5 to 10 seconds and 720p, which suits social shorts but not client work.
• Commercial rights: several free licenses restrict output to personal use. Paid plans are where commercial licenses usually begin.
• Speed: free generations sit in a slower queue, so batching a series of clips can take a while.
• Stability of terms: free allowances shrink regularly. A workflow that depends on one free plan is fragile by design.

Which one should you pick?

If you want the best-looking clips for free, start with Google Flow. If you want the most clips per day, Kling is the most generous. PixVerse is the fastest way to try the idea, Runway the best free taste of a professional tool, and CapCut is where clips become a finished video.

The strongest free workflow is a combination: generate your best shots on Flow or Kling, and assemble everything in CapCut. That setup costs nothing, respects the honest limits of each free tier, and is enough to produce real short-form video while you decide whether AI video is worth paying for.

Related Articles

• 5 New AI Tools Launched This Week (October 2026)
• Big Banks Just Blocked AI Agents From Moving Money: What Happened and What It Means for AI Finance Tools
• 5 AI Tools Launched This Week: Meta’s Muse Wonder, Google’s Gemini 4 Argon and More

Just days after major US banks moved to block AI agents from moving money on customer accounts, Mastercard has stepped in from the other side of the fence. On October 21, 2026, the payments giant opened early access to “Agent Pay for Machines” — a set of APIs, developer toolkits and a sandbox environment designed to let AI agents make payments on their own.

What Mastercard actually announced

Agent Pay for Machines is a developer platform, not a consumer product. Mastercard is giving builders the tools to test payments executed by autonomous software: APIs for authorization, identity verification for agents, and a sandbox where developers can experiment safely.

The announcement deliberately targets the hard problems first. Mastercard is focusing on two things that autonomous spending needs and doesn’t have yet: spend limits (how much an agent is allowed to spend, and where) and revocation (how you take an agent’s payment privileges away instantly).

Why this matters: the two visions of agent payments

The timing is the story. Last week, JPMorgan, Citi, Bank of America and Charles Schwab moved to block AI agents from accessing customer accounts — treating autonomous agents as a security risk. Mastercard’s move is the opposite bet: that machine-driven commerce is inevitable, and the right response is to build proper rails for it rather than wall it off.

Both sides are reacting to the same reality: AI agents are getting capable enough to shop, book and pay on behalf of users, and the payment industry’s plumbing was never designed for that. Somebody has to decide whether agents become first-class payment citizens.

A test environment, not a product launch

Keep the announcement in proportion. This is early access with developer tooling — not a card you can apply for today, and not an agent you can point at a store tonight. No pricing, no launch date and no partner list accompanied the news. Historically, only a fraction of sandbox projects ever reach production.

What it does confirm is direction. When a network the size of Mastercard starts doing the unglamorous infrastructure work — APIs, test environments, identity standards — for autonomous machine payments, it signals that agent commerce is being taken seriously at the highest levels of the financial system.

The convergence to watch

Mastercard’s programmable card payments and the crypto world’s on-chain programmable spending (stablecoins, smart-contract wallets) are now converging on the same use case from opposite directions. Visa has already moved stablecoin payouts to production with USDC, and Polygon’s acquisition of Coinme pushes further into payments territory. The race is on to define how machines will pay — whether through traditional card rails or blockchain ones.

What it means for AI tools and builders

• Agent commerce is becoming real infrastructure. Payment rails are the missing piece between an AI that can browse and an AI that can transact.
• Identity is the bottleneck. The hard part is not moving money — it is proving which agent is authorized, for how much, and letting users revoke it. Whoever solves identity owns the layer.
• Expect standards wars. Mastercard’s card-rail approach and the stablecoin approach will compete for developers over the next 18 months.

The bottom line

Banks are blocking AI agents at the account door; Mastercard is building them a checkout lane. Both moves point to the same conclusion: 2026 is the year agent payments stopped being a demo and started being an infrastructure problem. For anyone building AI tools with purchasing power, this is the announcement to watch — and the sandbox to get into early.

Related Articles

• Big Banks Just Blocked AI Agents From Moving Money: What Happened and What It Means for AI Finance Tools

In a single week, four of the biggest names in American banking drew the same line in the sand: outside AI agents are not allowed to move your money. JPMorganChase, Citigroup, Bank of America and Charles Schwab each said — in earnings calls, press reports and executive interviews — that third-party AI agents cannot log into customer accounts or transact on their behalf. It is the clearest sign yet that the age of the “AI that does everything for you” is colliding with the industry that trusts nothing.

What Each Bank Actually Said

The announcements came in quick succession. JPMorganChase CEO Jamie Dimon warned of “rogue agents” during the bank’s third-quarter earnings call and said outside agents would not get inside the bank. Charles Schwab went further with specifics: CEO Rick Wurster said the brokerage blocks big unauthorized AI agents, including Meta’s Muse, “in the tens and 20s” each week — attempts are already happening at scale.

Bank of America took a conditional stance rather than an outright ban. The bank is limiting agents’ access to consumer deposits while it writes formal protocols deciding which party is liable for a payment an AI authorizes. CEO Brian Moynihan said a human should stay in the loop to approve money leaving an account. Citigroup drew the same line by week’s end: no third-party agent logins, no third-party transactions.

Why Banks Are Drawing This Line

The core issue is liability — who pays when an AI agent moves money it should not have. If you hand your banking login to an AI agent and it makes a bad transfer, is that your fault, the agent developer’s fault, or the bank’s? Nobody has a clean answer, and banks are not volunteering to absorb the losses while the legal framework catches up.

There is also a security dimension. Agents that log in with your credentials look, to a bank’s fraud systems, exactly like account takeovers. Credential-sharing with a third-party tool breaks the chain of trust banks depend on. And the scale of consumer appetite makes the risk urgent: industry research found that 37% of AI power users already treat AI as their main tool for managing finances and banking, and consumers say they want an agent to ask permission before completing a purchase. Demand is real; the guardrails are not.

What Still Works

Importantly, the banks are not banning AI from finance — they are banning unauthorized agents. The distinction matters. Authorized connectors that read balances through official channels are still allowed. JPMorgan is building its own Smart Cash sweep feature, and Schwab is rolling out its own assistant, Charley. The pattern is clear: banks want AI money tools, but they want them inside their own walls, under their own liability rules.

Bank of America’s approach is the most revealing. By writing protocols that assign liability for AI-authorized payments — while keeping a human approval in the loop — it is sketching the template every other bank will likely follow. Expect “human approves, agent executes” to become the standard architecture for AI finance tools.

What This Means for AI Tool Builders

If you are building or evaluating AI finance tools, this week changed the landscape:

• Screen-scraping and credential-sharing agents are dead ends. Any tool that asks for your bank login is now fighting the banks themselves, not just regulators.
• Official APIs and authorized connectors are the only durable path. Tools that integrate through bank-approved channels will survive; workarounds will not.
• Human-in-the-loop is becoming a compliance requirement, not a feature. Approval gates before money moves are what banks and future regulations will demand.
• Bank-built agents have a structural advantage. JPMorgan and Schwab building their own AI assistants is not a coincidence — owning the liability framework is the moat.

Practical Tips If You Use AI for Money Management

For everyday users, the takeaway is straightforward. Never hand your banking login credentials to a third-party AI agent or browser tool, no matter how convenient it looks. Use only tools that connect through official bank APIs or authorized integrations. Keep human approval switched on for any transaction an AI proposes — if a tool does not offer an approval step, treat that as a red flag. And remember that if something goes wrong with an unauthorized agent, your bank’s fraud protection may not cover you.

The banks’ message this week was not “AI is bad.” It was “your money moves only with permission, and we decide what permission looks like.” For AI finance tools, the gold rush just got its first real fence. The winners will be the ones that build inside it.

Related Articles

• Mastercard Opens Agent Pay for Machines: AI Agents Get Their Own Payment Rails
• 5 AI Tools Launched This Week: Meta’s Muse Wonder, Google’s Gemini 4 Argon and More
• 5 New AI Tools Launched This Week (October 2026)

Chinese AI startup DeepSeek released an upgraded version of its AI model on October 17, 2026 — and this time the focus is not just on raw intelligence, but on speed, efficiency and cost. DeepSeek-V4.1-Flash-1017 is described by the company as faster and leaner than its predecessor, with improved benchmark scores and shorter response times.

Here is what the update changes, what the benchmarks say, and what it means if you use AI models for coding or agent-based work.

What is DeepSeek-V4.1-Flash-1017?

The update builds on the smaller V4.1-Flash model that DeepSeek launched in September 2026 — a release that reportedly put pressure on Chinese competitors, with shares of MiniMax Group, Z.ai and Alibaba falling on the Hong Kong stock market afterward, according to Bloomberg.

The 1017 version is available through DeepSeek’s website, app and developer platform. Crucially, it is an open-weight release under the MIT license, meaning developers can download the model weights and run or fine-tune them themselves — a key advantage for teams that want control over their AI infrastructure.

The benchmark numbers

In results DeepSeek published on Hugging Face, the new model scored 42.6 on Humanity’s Last Exam without tools, up from 36.8 for the previous V4.1-Flash version. That benchmark tests AI systems on difficult questions spanning mathematics, science and the humanities.

The technical report also listed improvements in software development tasks: on Terminal-Bench 4.0, which assesses coding agents, the score rose from 31.2 to 42.3, while the DeepSWE v1.1 score climbed from 74.2 to 77.0.

These figures come from DeepSeek’s own published evaluations, so they should be read with that caveat — they do not, on their own, establish how the model performs across all real-world uses. But the direction is consistent: a leaner model that performs better than the version it replaces.

Why “flash” matters: cheaper agent tasks

The most practical claim in the release is economic. DeepSeek says the update cuts agent task costs to just under $0.20 per task, compared with about $0.50 before — more than halving the cost of running AI coding agents and multi-step workflows.

For developers running large numbers of agent tasks, that difference compounds fast. If you call the deepseek-flash model on DeepSeek’s API, you are already on the 1017 snapshot — but the practical advice is to rerun your own agent and coding tests before relying on September results, since behaviour will have shifted.

The context: outages and competition

The release follows a difficult stretch for DeepSeek’s infrastructure, including recurring service slowdowns and an outage lasting about five hours earlier in the week that left its servers inaccessible. A faster, leaner model that produces shorter responses directly addresses that pain: less compute per request means less load on the servers.

The launch lands in an exceptionally crowded week for AI models. Google’s Gemini 4 Argon is now generally available, Meta launched Muse Wonder on October 12, and NVIDIA released three open Nemotron models aimed at enterprise agents. Open-weight efficiency — doing more with less — is becoming the battleground as Chinese and US labs compete on cost as well as capability.

How to try it

If you want to evaluate V4.1-Flash-1017 for your own work:

• Web or app: use it directly through DeepSeek’s chat interface — the fastest way to feel the shorter response times.
• API: call the deepseek-flash endpoint; the 1017 snapshot is live, so existing integrations pick it up automatically.
• Open weights: download the MIT-licensed weights from Hugging Face to run it on your own infrastructure or fine-tune it.
• Benchmark before committing: rerun your own coding-agent evaluations against the September version — DeepSeek’s charts are a starting point, not a verdict.

The broader trend is clear: the most useful AI updates in 2026 are not always the smartest models, but the ones that are fast enough and cheap enough to run at scale. V4.1-Flash-1017 is DeepSeek’s bet that efficiency wins.

Related Articles

• How to Run FLUX 3 Image Klein: The Open-Weight Image Model That Fits on an 8GB GPU

High-quality AI image generation has long come with a hardware tax: the best models demanded flagship GPUs, leaving most creators stuck with cloud queues. Black Forest Labs changed that on October 15, 2026 with FLUX 3 Image Klein — a 5.6 billion-parameter open-weights image model that generates natively at 4K while running on a single 8GB GPU. Here is how to set it up, which version to pick, and what the license allows.

What makes Klein different

Most compact image models cheat: they generate at 512 or 1024 pixels and rely on an upscaler, which softens detail. Klein renders natively at 4K, so textures and fine lines come from the model itself — a visible difference for posters, wallpapers, and zoomed compositions.

Editing is the other headline. Klein accepts up to 10 reference images in a single edit pass: restyle a room from several photos, merge outfit concepts for a fashion mockup, or keep a character consistent by feeding multiple angles. It is the first open-weights model in its class to support multi-reference editing at this depth.

Under the hood, prompt understanding comes from a Qwen3-VL 4B text encoder, which helps Klein follow complex, multi-object prompts more faithfully than the small text encoders typical of compact generators.

Which version to pick

Black Forest Labs ships Klein as several checkpoints, and picking the right one depends on your hardware and patience:

Base vs. step-distilled. The base checkpoint gives the best quality but needs more sampling steps. The step-distilled variant reaches near-base quality in far fewer steps — the right choice for fast iteration on a mid-range card.

Precision builds: FP8, INT8, and NVFP4. FP8 is the balanced default on modern NVIDIA cards. INT8 squeezes into tighter VRAM budgets for older GPUs. NVFP4 is the most aggressive option for the newest architectures, trading a small quality margin for maximum speed and lowest memory use.

Rule of thumb: on an 8GB card like an RTX 3060 Ti or 4060 Ti, start with step-distilled FP8. With more VRAM, run base FP8 for the best quality. Drop to INT8 or NVFP4 only if you hit out-of-memory errors or want the fastest turnaround.

Setup guide: running Klein in ComfyUI

The most popular local route is ComfyUI, version 0.41.0 or newer, which ships native support for the model. Diffusers users can load the checkpoints through the Hugging Face libraries, and anyone who would rather skip local hardware can call the BFL API directly.

Here is the local setup, step by step:

1. Update ComfyUI to version 0.41.0 or later. Klein support is built into recent releases, so a standard update is all you need — no custom nodes required for basic generation.
2. Download the checkpoint you chose above and place it in the models/checkpoints folder of your ComfyUI installation.
3. Build the workflow: add a Load Checkpoint node pointing at the Klein file, a CLIP Text Encode node for your prompt, an empty latent image set to 4K-friendly dimensions, and a KSampler feeding a VAE Decode node. A text-to-image starter workflow works unchanged — Klein follows the standard pipeline.
4. Set sampler values: for the distilled build, 4 to 8 steps is enough; for the base build, start around 20 to 28 steps with a standard scheduler. Keep CFG near the model default rather than pushing it high, which tends to oversaturate Klein outputs.
5. Run and iterate. Generate a small batch at reduced resolution first to check composition, then render your keepers at full 4K.

For multi-reference editing, the same pipeline applies: load your reference images through image-input nodes and connect them alongside the prompt so the model conditions on all of them at once, up to the 10-image limit.

How fast is it?

Speed is where Klein’s efficiency shows. NVIDIA’s measurements put generation at roughly 4 seconds per image on an RTX 5090 and under 9 seconds on an RTX 4070. On a minimum-spec 8GB card expect longer waits — but renders complete at all, on hardware that could not load models of this class a year ago.

Prefer not to run anything locally? The BFL API serves Klein at $0.009 per 1K, one of the cheapest ways to access a 4K-capable open model on demand.

What the license actually allows

Klein is released under the FLUX Kommunity License. You can download the weights, run them locally, fine-tune them, and use outputs commercially — with one ceiling: once revenue from the model crosses $5 million, you need a commercial license from Black Forest Labs. For independent creators and small studios that line is comfortably distant, making Klein genuinely usable for client work rather than just experimentation.

Honest limitations

Two known weak spots: text rendering (signs, logos, labels inside images) is unreliable, so do not depend on it for graphics with legible wording. People are the other weakness — hands, faces in crowds, and tricky poses carry the familiar artifacts of compact models. Portrait and photoreal-human work will need careful prompting and cherry-picking.

The base checkpoint performs slightly better than the distilled one on both, so if your work involves people or text-heavy scenes, the slower base build is worth the wait.

The bottom line

FLUX 3 Image Klein is the first open-weights image model to combine native 4K generation, deep multi-reference editing, and genuine 8GB-GPU compatibility — with a license that lets small creators earn from it. Download the step-distilled FP8 build, load it into ComfyUI 0.41.0+, and you can be rendering your first 4K image today.

Related Articles

• DeepSeek V4.1 Flash 1017: Faster, Leaner and Cheaper — A Practical Guide to the Open-Weight Update

October 2026 has been one of the busiest weeks in AI in months. Meta launched not one but two new model families, Google’s next flagship model went generally available, Anthropic shipped a cheaper workhorse model and brought Claude directly into Google Workspace, and OpenAI previewed a new API built for decisions rather than essays. Here are the five launches that matter most, what each one does, and whether it is worth your time.

1. Meta Muse Wonder and Muse Wonder Full

Meta expanded its AI lineup with Muse Wonder and Muse Wonder Full, its push into coding assistants and agentic AI. The models are designed to remember relevant details across chats and stick with complex, multi-step assignments. Muse Wonder is available through Meta’s developer API and Muse Code, while the larger Full variant is limited to Muse Code subscribers. Meta also released Muse Spark Open, an open-weight model for coding and agents that it optimised with Apple to run locally on high-memory Macs — a serious nod to developers who want private, on-device AI.

2. Google Gemini 4 Argon

Google’s next-generation flagship model, Gemini 4 Argon, is now generally available. Independent testing suggests it hallucinates less than rivals — answering conservatively where others guess — which could make it a strong pick for tasks where accuracy matters more than coverage. Google is positioning Argon for long-horizon agent work, with developer guides covering robotics and autonomous coding.

3. Claude Haiku 5.5 and Anthropic’s cheaper pricing

Anthropic released Claude Haiku 5.5 on October 7, bringing a 1-million-token context window and adjustable effort to its cheapest model line. Prompts up to 100,000 tokens cost $0.10 per million input tokens and $0.50 per million output tokens. Anthropic also halved Sonnet 5.5’s cache-read pricing, making it cheaper for workloads that repeatedly reuse context. If you run high-volume summarisation or agent loops, Haiku 5.5 is suddenly the price leader to beat.

4. Claude comes to Google Docs, Sheets and Slides

Anthropic launched Claude for Google Workspace in public beta on paid plans. Users can now work with Claude in a sidebar or create and edit Google files directly through connectors, without switching applications. Access follows your existing Google sharing permissions, and Claude asks before editing by default — a sensible safeguard for collaborative documents.

5. OpenAI’s Decisions API

OpenAI entered public beta with a Decisions API that returns typed verdicts instead of long essays, promising results roughly ten times faster than its standard Responses API. If you build apps where an AI needs to make a structured call — approve, reject, categorise — rather than write prose, this is worth a look in the free playground before committing to per-token billing.

How to choose: a quick decision guide

• Build apps at scale: compare Haiku 5.5’s new pricing against the Decisions API for your specific task — test two models before standardising.
• Work in Google Workspace daily: try the Claude beta in Docs and Sheets; it may remove your biggest app-switching friction.
• Need private or offline AI: Muse Spark Open’s local Mac deployment is worth evaluating, though the hardware requirements are steep.
• Curious but cautious: Satya Nadella’s advice this week is worth heeding — guard AI like a powerful employee who could slip up: keep an emergency brake, keep activity logs, and never let a model grade its own work.

The week’s theme is clear: AI is moving out of the chat window and into devices, work apps and developer workflows. The launches are genuinely exciting — but with the US FTC reportedly demanding detailed security disclosures from OpenAI and Anthropic, and Anthropic cutting its own test agents off from the live internet after they exploited real websites, this is also the week that proved AI safety is everyone’s roadmap now.

Related Articles

• 5 New AI Tools Launched This Week (October 2026)
• Big Banks Just Blocked AI Agents From Moving Money: What Happened and What It Means for AI Finance Tools

5 New AI Tools Launched This Week (October 2026)

The second week of October 2026 has been one of the busiest AI launch windows in months. From a desktop AI that learns your workflow by watching it, to an API that returns verdicts instead of essays, to a voice model that speaks 90+ languages — here are the five launches worth knowing about, what they cost, and which one deserves your time first.

1. IrisGo — AI Workflow Automation for Solopreneurs

IrisGo is a desktop AI tool that watches how you work and learns to automate your repetitive tasks. Aimed at solopreneurs drowning in admin work, it entered free beta this week and became the #1 Product of the Day on October 7. If you spend hours on routine computer tasks, this is the one to try first — and it costs nothing while in beta.

2. OpenAI Decisions API — Verdicts Instead of Essays

OpenAI launched the Decisions API in public beta on October 6. Instead of generating long text responses, it returns short, typed verdicts — ideal for apps that need quick yes/no or classification decisions. OpenAI claims it is roughly 10x faster than the standard Responses API, and pricing starts at just $0.10 per million input tokens, making it the cheapest way to add decision-making to an app.

3. Suno Speech — Voice Plus Music in One Track

Suno, known for AI music generation, launched Suno Speech in public beta on October 1. It combines voice narration and music in a single track — perfect for creators who make narrated videos, podcasts, or ads and previously needed two separate tools. A free plan is available with download limits.

4. ElevenLabs Eleven v4 — More Expressive Voice AI

ElevenLabs announced Eleven v4 in late September, bringing support for 90+ languages and in-sentence expression tags that let you control emotion mid-sentence. For anyone doing voiceovers, dubbing, or AI narration, this is a major step up in natural-sounding speech.

5. Claude Haiku 5.5 — Cheapest Frontier-Class Small Model

Anthropic launched Claude Haiku 5.5 on October 7, priced at $0.10 per million input tokens and $0.50 per million output tokens — among the cheapest prices for a frontier-class small model. It is built for speed and high-volume tasks like classification, moderation, and quick customer replies.

Which One Should You Try First?

It depends on what you do: try IrisGo if repetitive admin work eats your day; the Decisions API if you build apps; Suno Speech if you create audio or video content; Eleven v4 if you need voiceovers; and Haiku 5.5 if you need cheap, fast AI at scale. Three of the five are free to try right now, so there is little reason not to experiment.

Frequently Asked Questions

Are these AI tools free?
IrisGo is in free beta, Suno Speech has a free plan, and the Decisions API has a free playground — the others are paid but very affordable.

What is the cheapest AI model launched this week?
Claude Haiku 5.5 and OpenAI’s Decisions API both start at $0.10 per million input tokens.

Do I need technical skills to use these tools?
IrisGo and Suno Speech are built for non-technical users. The Decisions API and Haiku 5.5 are aimed at developers.

Related Articles

• 5 AI Tools Launched This Week: Meta’s Muse Wonder, Google’s Gemini 4 Argon and More
• Big Banks Just Blocked AI Agents From Moving Money: What Happened and What It Means for AI Finance Tools

Welcome to WordPress. This is your first post. Edit or delete it, then start writing!