Quick answer
Muse Spark Open is Meta’s new open-weight AI model, which Meta calls the most capable US open-weight model to date. It is natively multimodal, with 816 billion total parameters (42 billion active) and a 1 million token context window, and it is aimed at coding, AI agents and long-running tasks. You can download the weights or use it through hosted platforms such as OpenRouter. Running it at home needs very expensive hardware. In early tests it was strong for an open model but not the best overall. The same week brought unconfirmed sightings of Claude checkpoints that may be Fable 5.5 and Opus 6, Opus 5.5 Fast Mode leaving research preview, and a Claude Code /export command.
Key takeaways
- Muse Spark Open: 816B total / 42B active parameters, 1M-token context, text, image, video and audio input, open weights on Hugging Face (gated)
- Tests: it built the most playable game in one four-model open-weight test (about $0.59 and 20 minutes), but sat just behind Qwen 3.8 27B on the creator’s own benchmark
- Anthropic rumours: food-themed early-access checkpoints may be Fable 5.5 and Opus 6. Nothing is confirmed
- Shipped: Opus 5.5 Fast Mode is out of research preview at $8 / $40 per million input/output tokens, and Claude Code can export a session to a text file
- Also: OpenAI capacity fixes and a voice upgrade, Microsoft’s MAI Base 2, free DeepSeek V4.1 Flash and Nemotron 3.5 Super VL in OpenCode, and a new Flux 3 image model
In this guide
- Watch: the source video
- What is Muse Spark Open?
- Can you run Muse Spark Open locally?
- How did Muse Spark Open do in hands-on tests?
- Is Claude Opus 6 or Fable 5.5 about to launch?
- What changed in Opus 5.5 Fast Mode and Claude Code?
- What is OpenAI shipping this week?
- What is Microsoft’s MAI Base 2?
- Which free and open models are new?
- What else happened: Gemini 4 Argon and a flying Tesla
- What should small businesses do with this news?
- Frequently asked questions
- The bottom line
Meta has dropped another big model. Muse Spark Open is an open-weight model you can, in theory, run on your own machines, and it arrives in a week packed with leaks, price changes and free models.
Here is what was announced, what is only rumoured, and what is actually worth your time if you run a small business or build automations. If you missed last week’s round-up, start with Fable 5.5 Release Rumours, OpenAI Spy Arrest and a Big Claude Code Update.
🛠️ Practical guides
Want to act on this week’s news? These step-by-step guides answer two follow-up questions:
- Should You Run Muse Spark Open Locally or Use It Through an API? — the hardware reality, a rough memory calculator and a simple way to decide between local and hosted use.
- How to Export a Claude Code Session and Reuse It in Another AI Agent — how to use
/export, clean the file and hand the work to Codex or another agent, or turn it into a reusable skill.
Watch: the source video
What this video covers: World of AI (@intheworldofai) covers Meta’s Muse Spark Open release and runs it through coding and 3D tests. It also covers reported Claude early-access checkpoints linked to Fable 5.5 and Opus 6, a new GPT-6.1 Astra snapshot, Microsoft’s MAI Base 2, Opus 5.5 Fast Mode, Claude Code’s export command, OpenAI’s shipping-pledge updates, Nvidia’s Nemotron 3.5 Super VL, Flux 3, Gemini 4 Argon in Antigravity and Tesla’s airborne Roadster demo.
Video: World of AI (@intheworldofai). This summary and the opinions in it are our own. Leaks and rumours are labelled as unconfirmed.
What is Muse Spark Open?
Short answer: it is Meta’s biggest open-weight release so far, built for heavy work rather than quick chat. Here are the headline specs Meta has shared:
| Spec | Muse Spark Open |
|---|---|
| Total parameters | 816 billion |
| Active parameters per token | 42 billion |
| Context window | 1 million tokens |
| Inputs | Text, images, video and audio (natively multimodal) |
| Built for | Coding, AI agents, complex software engineering and long-horizon reasoning |
| Where to get it | Open weights on Hugging Face (gated), plus hosted platforms such as OpenRouter |
“Active parameters” matters because only part of the model works on each token. That is why a model this large can run faster than its size suggests. It still needs enough memory to hold the whole thing.
Meta also showed the model working on its own: making changes to an iOS app and controlling DaVinci Resolve to edit video footage. The creator believes it is based on Muse Spark 1.2, which is no longer the strongest model around, so expect it to trail the very top closed models.
⚠️ Warning: The weights are gated. To download them on Hugging Face you have to give your full name and date of birth. Read the licence there before you use the model in client or commercial work.
Can you run Muse Spark Open locally?
Yes, but “locally” here means serious hardware. The setups mentioned in the video were:
- Two Apple Mac Studios with M5 Ultra chips, which the video says is enough to run the full model on a desktop.
- An M3 Ultra Mac Studio with about 512 GB of unified memory, running the model through llama.cpp, where it reportedly worked surprisingly well.
- Three Nvidia DGX Spark units with a 2-bit quantised version, which managed about 7 tokens per second.
For most small businesses, a hosted API is the realistic way to try it. You pay per use, and you can switch away if it doesn’t suit you. Our guide Should You Run Muse Spark Open Locally or Use It Through an API? has a rough memory calculator and a five-question checklist to help you decide.
How did Muse Spark Open do in hands-on tests?
On the benchmarks shared at launch it does well on agentic, knowledge-work and coding tasks, but it doesn’t beat every Chinese frontier model, such as Qwen 3.8. On the creator’s own World of AI benchmark it landed a couple of points behind Qwen 3.8 27B, a much smaller model. His view of its front-end design taste was “pretty good” for a local model, though not the best.
Three hands-on tests stood out:
| Test | What happened | Takeaway |
|---|---|---|
| Pixel “survivors” game in HTML canvas, against three other open-weight models | Muse Spark Open built the most playable game, with enemy hordes, damage numbers, a level-up system and zero crashes, for about $0.59 in about 20 minutes. Inkling was fastest (11 minutes) with the best visuals, but its weapons did no damage. HY4 Preview took 52 minutes. Step 5 Preview crashed and needed a fix | Best working result, not fastest or cheapest |
| Voxel pagoda (3D spatial reasoning) | A five-tier pagoda with a gate on the first attempt, but cropped and short on detail. It took 37 minutes and over 15k tokens at about 7 tokens/sec on three DGX Sparks (2-bit quant) | Functional code, weaker visual composition |
| Voxel city blown up with TNT, against Muse Wonder | Muse Wonder made the bigger, better explosion. Muse Spark Open finished faster at roughly 10 times lower cost | Cheaper and quicker, lower quality |
💡 Tip: Treat any single test as one data point. Before you rely on a model, run it on three or four of your own real tasks. For more on this, see our guide on how to verify AI model claims.
Like practical takes on AI news? Each week we turn the headlines into clear next steps for small businesses. Browse the latest guides on Prodify Digital.
Is Claude Opus 6 or Fable 5.5 about to launch?
Short answer: maybe, but there is no announcement. The creator expected Fable 5.5 this week and now doubts it will arrive in time. What has appeared instead is a set of early-access (EAP) checkpoints on Artificial Analysis’s website.
🔎 Unconfirmed / rumour: Artificial Analysis has a history of listing unreleased Claude models under food code names. Two earlier ones matched real launches: Claude Beignets EAP was linked to Sonnet 5.5, and another was linked to Haiku 5.5. Two newer code names, reported as Oniiri EAP and Gabagool EAP, briefly appeared with settings that look like Fable- and Opus-tier models, then were removed. The guess is that Oniiri is Fable 5.5 and Gabagool is Opus 6 or an updated Opus checkpoint. Anthropic has not confirmed any of this.
| Signal | What was reported | How much weight to give it |
|---|---|---|
| Oniiri EAP | Four reasoning levels (low, medium, high, extra high). Reportedly falls back to Opus 4.8, like earlier Fable and Opus-tier models | Interesting pattern, not proof |
| Gabagool EAP | Fable/Opus-tier settings. Demos show strong 3D visuals, tool use and Python coding with Blender | Impressive demos, origin unverified |
| “Claude Next” silent routing | Some users may be routed to a newer model in Claude Code and on the web when Opus is selected | Hard to check from the outside |
| Short AI-made film | A short film credited to the Opus EAP checkpoint | A teaser, not a benchmark |
If Anthropic shipped Fable 5.5 and then Opus 6 before its expected IPO, that would shake up the AI race. Until then, our guide on how to verify AI model leak rumours explains which of these signals actually mean something.
What changed in Opus 5.5 Fast Mode and Claude Code?
Two practical changes from Anthropic actually shipped this week.
Opus 5.5 Fast Mode is out of research preview
Fast Mode makes Claude Opus 5.5 up to 2.5 times faster, and the waitlist has gone. Speed costs extra: $8 per million input tokens and $40 per million output tokens. It is also available through Microsoft Foundry on Azure, with Google Cloud support coming soon. In Claude Code, Fast Mode is unchanged and still uses extra usage credits.
💡 Tip: Save Fast Mode for jobs where waiting really costs you, such as a live client demo or a long agent run you are supervising. For overnight or batch work, normal speed is usually fine and cheaper.
Claude Code can now export a session
The new /export command, followed by a file name, saves your whole Claude Code session as a plain text file. That sounds small, but it is handy. You can pass the file to another AI agent to carry on where you left off, or review it to spot repeated steps worth turning into a custom skill.
Our step-by-step guide, How to Export a Claude Code Session and Reuse It in Another AI Agent, covers cleaning the file before you share it and gives you a copy-and-paste handoff prompt.
What is OpenAI shipping this week?
OpenAI is now on day 11 of its 28-day shipping pledge. The useful bits:
- Day 10: a usage-limit reset for everyone, given because OpenAI missed shipping a new feature that day.
- Day 11: changes to cut the “selected model is at capacity” errors, including more compute. If you do hit the error, your session should now resume automatically instead of getting stuck.
- Voice mode upgrade: a new GPT Live 1.1 model across ChatGPT, Codex and Dots, with more voices in French, German, Hindi, Mexican Spanish, Korean and Japanese. The “checking” message that sometimes got stuck during voice sessions should be fixed. API support is said to be coming soon.
🔎 Unconfirmed / rumour: A newer internal snapshot of GPT-6.1 Astra has reportedly been found. The creator says OpenAI shelved an earlier version when the latest Opus launched and has since trained it further, so it could ship within the pledge period, perhaps in the next couple of weeks. OpenAI has not announced it.
⚠️ Warning: An open-source project called “Chat on Steroids” claims to turn ChatGPT into a Codex-style coding environment that uses your ChatGPT limits instead of your Codex limits. The “practically unlimited GPT-6” claim is unconfirmed. Unofficial tools like this can also break a service’s terms, so check before you connect a work account.
What is Microsoft’s MAI Base 2?
Microsoft is reportedly finishing pre-training its next frontier model, MAI Base 2. The reported figures:
- 2.8 trillion total parameters, with 80 billion active.
- Trained on 16,000 Nvidia GB200 GPUs in just 38 days.
- Multimodal, understanding text and images.
- Microsoft claims 4 to 12 times more efficient pre-training than DeepSeek V4 Pro Base. Early comparisons reportedly show it ahead of nearly every Chinese base model tested.
Keep one thing in mind: this is a base model. It hasn’t had the post-training that makes models good at following instructions and doing real tasks, so the final product could look quite different. It is still notable that Microsoft built it entirely in-house rather than relying on a partner.
Which free and open models are new?
| Model | What it is | Where you can try it |
|---|---|---|
| DeepSeek V4.1 Flash | A fast model with a 200K-token context window | Free in the OpenCode coding agent, reportedly for this week |
| Nemotron 3.5 Super VL (Nvidia) | An open model for agents: 120B total / 12B active. Reads text, images, documents, charts and video in one model. Nvidia says it finishes coding tasks up to four times faster while keeping strong SWE-bench Pro accuracy, and stays quick even with 100K-token inputs | Free in OpenCode |
| Flux 3 image model (Black Forest Labs) | An open-weight image model with 5.6B parameters, native 4K output, pixel-precise editing and bounding-box control. It can be fine-tuned, and comes in FP8, NVFP4 and INT8 versions | Runs on a single consumer GPU. A distilled 4K version is coming to the API |
⚠️ Warning: Free tiers are a great way to test, but they can change or end without notice. Don’t build a client workflow that depends on a free model staying free, and test on dummy data first.
What else happened: Gemini 4 Argon and a flying Tesla
- Gemini 4 Argon in Antigravity. Google is rolling Argon out to paid Google AI Pro subscribers in its Antigravity coding tool as it adds capacity. Argon gets its own weekly usage quota, so it won’t eat into your Gemini 3.8 Flash allowance. The creator still prefers Codex and Claude Code for coding.
- Tesla’s airborne Roadster. At the next-generation Roadster unveiling, hosted by Tesla and SpaceXSI, the car reportedly jumped over a row of Teslas and landed using compressed-air thrusters developed with SpaceXSI. It was also shown driving upside down using suction. The creator links it to AI-driven real-time control. It is a stunt rather than something practical, and the exact role of AI wasn’t detailed.
What should small businesses do with this news?
| Story | Status | What we would do |
|---|---|---|
| Muse Spark Open | Released | Try it through a hosted API on one real task before thinking about hardware |
| Opus 5.5 Fast Mode | Released | Use it only where speed saves real time or money |
Claude Code /export | Released | Export long sessions and reuse them for handoffs or new skills |
| Fable 5.5 / Opus 6 checkpoints | Rumour | Prepare a test set. Switch nothing until launch |
| GPT-6.1 Astra snapshot | Rumour | Wait for an announcement |
| MAI Base 2 | Reported, pre-training only | Background only. No action needed |
| Free DeepSeek and Nemotron in OpenCode | Available now, may change | Test on dummy data while it’s free |
The pattern for small businesses is the same each week. Act on what has shipped, test it on your own work, and keep rumours in a “watch” column. If you are juggling several models already, our guide to AI Model Routing for Small Business: Match Each Task to the Right Model helps you match each task to the right one.
Want to turn AI news into a working system? Start with one task, one model and one measure of success. Then decide whether to run Muse Spark Open locally or set up session exports in Claude Code, using our step-by-step guides.
Frequently asked questions
What is Muse Spark Open?
It is Meta’s open-weight AI model, with 816 billion total parameters, 42 billion active parameters and a 1 million token context window. It accepts text, images, video and audio, and it is aimed at coding, AI agents and long-running reasoning tasks.
Is Muse Spark Open free?
The weights can be downloaded from Hugging Face after a gated sign-up, which asks for your full name and date of birth. Running it yourself needs expensive hardware. Hosted platforms such as OpenRouter charge per use. In one test, a complete game cost about $0.59 through a hosted provider.
Can I run Muse Spark Open on a normal laptop?
Not realistically. The setups shown used two M5 Ultra Mac Studios, an M3 Ultra Mac Studio with about 512 GB of memory, or three Nvidia DGX Sparks running a 2-bit version at about 7 tokens per second.
Is Claude Opus 6 out?
No. As of 16 October 2026, early-access checkpoints reported as Oniiri and Gabagool have been spotted and linked to Fable 5.5 and Opus 6, but Anthropic has not announced either model.
How much does Opus 5.5 Fast Mode cost?
$8 per million input tokens and $40 per million output tokens, in return for up to 2.5 times faster responses. In Claude Code it uses extra usage credits.
How do I export a Claude Code session?
Type /export followed by a file name in Claude Code. It saves the session history as a plain text file. Our guide explains how to clean it and reuse it with another agent.
The bottom line
Muse Spark Open is a real step forward for open-weight AI, and it is genuinely useful that a model this capable can run outside a big lab’s cloud. For most small businesses, though, the smart move is to try it through a hosted API on one real task first. The Claude Opus 6 and Fable 5.5 talk is still rumour. The changes you can use today are Fast Mode, where speed matters, and Claude Code’s new export command for handing work between agents.
Related posts
- Fable 5.5 Release Rumours, OpenAI Spy Arrest and a Big Claude Code Update
- How to Verify AI Model Leak Rumours Before You Switch Tools
- AI Model Routing for Small Business: Match Each Task to the Right Model
[…] Yes. Weights and hosted APIs. See our Muse Spark Open breakdown […]