We’re excited to announce that starting today, Microsoft Foundry customers can access high performance, low latency inference performance of popular open models hosted on the Fireworks cloud from their Foundry projects, and even deploy their own customized versions, too!

As part of the Public Preview launch, we’re offering the most popular open models for serverless inference in both pay-per-token (US Data Zone) and provisioned throughput (Global Provisioned Managed) deployments. This includes:

• Minimax M2.5 🆕
• OpenAI’s gpt-oss-120b
• MoonshotAI’s Kimi-K2.5
• DeepSeek-v3.2

For customers that have been looking for a path to production with models they’ve post-trained, you can now import your own fine-tuned versions of popular open models and deploy them at production scale with Fireworks AI on Microsoft Foundry.

 

The Microsoft Foundry model catalog showing the new Fireworks on Foundry model collection.

Serverless (pay-per-token)

For customers wanting per-token pricing, we’re launching with Data Zone Standard in the United States. You can make model deployments for Foundry resources in the following regions:

• East US
• East US 2
• Central US
• North Central US
• West US
• West US 3

Depending on your Azure subscription type, you’ll automatically receive either a 250K or 25K tokens per minute (TPM) quota limit per region and model. (Azure Student and Trial subscriptions will not receive quota at this time.)

Per-token pricing rates include input, cached input, and output tokens priced per million tokens.

 

 

As we work together with Fireworks to launch the latest OSS models, the supported models will evolve as popular research labs push the frontier!

 

Provisioned Throughput

For customers looking to shift or scale production workloads on these models, we’re launching with support for Global provisioned throughput. (Data Zone support will be coming soon!)

Provisioned throughput for Fireworks models works just like it does for Foundry models: PTUs are designed to deliver consistent performance in terms of time between token latency. Your existing quota for Global PTUs works as does any reservation commitments!

 

^ Calculated as p50 request latency on a per 5 minute basis.

 

Custom Models

Have you post-trained a model like gpt-oss-120b for your particular use case? With Fireworks on Foundry you can deploy, govern, and scale your custom models all within your Foundry project. This means full fine-tuned versions of models from the following families can be imported and deployed as part of preview:

• Qwen3-14B
• OpenAI gpt-oss-120b
• Kimi K2 and K2.5
• DeepSeek v3.1 and v3.2

 

The new Custom Models page in the Models experience lets you initiate the import process for copying your model weights into your Foundry project.

 

Importing Custom Models into Microsoft Foundry is available under Build -> Models -> Custom Models.

For performing a high-speed transfer of the files into Foundry, we’ve added a new feature to Azure Developer CLI (azd) for facilitating the transfer of a directory of model weights. The Foundry UI will give you cli arguments to copy and paste for quickly running azd ai models create pointed to your Foundry project.

Enabling Fireworks AI on Microsoft Foundry in your Subscription

While in preview, customers must opt-in to integrate their Microsoft Foundry resources with the Fireworks inference cloud to perform model deployments and send inference requests. Opt-in is self-service and available in the Preview features panel within your Azure portal.

 

The Azure Preview features panel for an Azure subscription where you can enable the Fireworks on Foundry experience.

For additional details on finding and enabling the preview feature, please see the new product documentation for Fireworks on Foundry.

 

Frequently Asked Questions

How are Fireworks AI on Microsoft Foundry models different than Foundry Models?

Models provided direct from Azure include some open-source models as well as proprietary models from labs like Black Forest Labs, Cohere, and xAI, and others. These models undergo rigorous model safety and risks assessments based on Microsoft’s Responsible AI standard.

For customers needing the latest open-source models from emerging frontier labs, break-neck speed, or the ability to deploy their own post-trained custom models, Fireworks delivers best-in-class inference performance. Whether you’re focused on minimizing latency or just staying ahead of the trends, Fireworks AI on Microsoft Foundry gives you additional choice in the model catalog.

Still need to quantify model safety and risk? Foundry provides a suite of observability tools with built-in risk and safety evaluators, letting you build AI systems confidently.

 

How is model retirement handled?

Customers using serverless per-token offers of models via Fireworks on Foundry will receive notice no less than 30 days before potential model retirement. You’ll be recommended to upgrade to either an equivalent, longer-term supported Azure Direct model or a newer model provided by Fireworks.

For customers looking to use models beyond the retirement period, they may do so via Provisioned throughput deployments.

 

How can I get more quota?

For TPM quota, you may submit requests via our current Fireworks on Foundry quota form.

For PTU quota, please contact your Microsoft account team.

 

Can you support my custom model?

Let’s talk! In general, if your model meets Fireworks’ current requirements, we have you covered. You can either reach out to your Microsoft account team or your contacts you may already have with Fireworks.

 

In the computer playground wars of the 1980s, children would spend hours extolling the virtues of either the ZX Spectrum or Commodore 64, depending on which side of the fence they sat. As the arguments raged, those who owned an Amstrad CPC machine tended to watch from the sidelines; but, deep down, they knew their chosen machine could very much hold its own.

As time has gone on, there’s been a growing appreciation for the 8-bit computers created by Lord Sugar’s company. In recent years, a small but nonetheless loyal community has been creating a string of games that push the machines to their limits (check out Pinball Dreams as proof). They’ve also been producing hardware to take the CPC to the next level. A good example of this is the M4 board, which not only enables wireless LAN but also allows an SD card to be used for storage.

Joining the hardware roster is the PicoCPC, a new multi-purpose add-on that can be used across the CPC range, including the Plus machines launched in 1990. It’s impossible to sum up its capabilities in one sentence. Rather, this neat little device — built around a Raspberry Pi Pico 2 — provides a whole host of benefits for machines that maxed out at 128kB of memory. 

For starters, the PicoCPC extends the memory up to 1024kB. It also emulates a floppy disk controller, allows the use of up to 16 emulated ROMs (for instant access to the likes of the Protext word processor and alternative operating systems such as SymbOS and FutureOS), and enables cartridge-loading of software produced for the GX4000 console. It adds six-voice audio, courtesy of the PlayCity sound card emulation, and there’s also a clock. For CPC enthusiasts, it’s fast becoming an essential upgrade.

Key to success

The idea emerged after Stéphane Plantard noticed a problem with many previous CPC add-ons. “I discovered there were a lot of expansions for Amstrad’s early computers, but they were expensive, rare, and made with deprecated chips,” he says. Over time, he became very familiar with the inner workings of the CPC. He produced an external Gotek drive for the CPC 664 and the CPC 6128, as well as a device that powers the CPC and allows it to be connected to a TV instead of the bundled monitor.

“Then my friend Freddy started a project called PicoMEM for old PCs,” he says, of a device based on the RP2040 microcontroller that runs emulated 8-bit ISA boards on a real PC. “I thought I could do a similar card for the CPC, so I started to write some CPC-related code on the PicoMEM to test the SD card readings, which allowed me to become familiar with Raspberry Pi Pico. Freddy then pushed me to create my own card, and I’ve produced three prototypes so far.”

Stéphane says he approached the project according to the agile principles, which essentially means he sought to be flexible, open to change, and willing to work closely with the CPC community (there are twelve principles in total, and they’re part of the Agile Manifesto published in 2001). “It means I maintain a backlog of ideas and set up my epics and my sprints,” he explains. “Having a strategy is a requirement when you work alone on such a big project.”

In essence, Stéphane has been working in short development cycle bursts, or “sprints”, that tend to stretch to a week. “It has allowed me to achieve some progress and reach goals; it’s often enough to stay motivated,” he notes. Given the project has taken a year so far, motivation has proven important. Stéphane has certainly seen the many benefits of basing the project around a Raspberry Pi Pico 2, and he says it’s preferable to the alternatives he considered.

“The CPC generates a lot of signals that the card has to act upon, so an STM32 microcontroller would not be fast enough,” he says, of one potential choice. He could have opted for an FPGA solution instead, but this, too, posed problems: “It would have required expensive chips, and I don’t know how to work with them,” he adds. “At the end of the day, the Pico 2 is cheap and fast. Its two cores and the PIOs make it a lot more suitable for this kind of task than any other available microcontroller. It uses one core to manage CPC I/O, a second core to run the emulations, and the PIO pilots the multiplexors to get addresses and pull/push data from/to the CPC.”

Quick upgrade

Some of Stéphane’s ideas are still being worked on. “The PicoCPC does not support USB mice and joysticks yet. It’s in the backlog but not done,” he says. But, as it stands, the PicoCPC makes a big difference to the experience of using an Amstrad CPC, and it also addresses a few practical issues.

“An ordinary CPC user would want to use an original computer to play old games from back in the day,” Stéphane says. Since the tape-based CPC 464 is easier to find than the disk-based CPC 6128, the PicoCPC can be used to transform a 464 into a 6128 by adding an emulated floppy drive, more RAM, a later version of BASIC (v1.1), and the C3 mode which allows access to the extra RAM beyond the base 64kB. “All the CPC 6128 games and demos will then work on a CPC 464.”

Likewise, on a CPC 6128, PicoCPC can override the internal floppy. “Users won’t have to take care of deteriorated, old floppy disks,” Stéphane says. But that’s just touching the surface, he adds. “Many new games developed in recent years require 512kB of memory,” he explains. “The Pico does not offer enough memory for that, which is why I added an SRAM chip on the card.”

He hopes the availability of PlayCity will push future software developers to use the expanded audio capabilities, and he’s excited about other possibilities. “I plan to add proper hard-disk emulation, AdLib music card emulation, and new blank floppy disk creation,” he reveals. “But the first plan for the PicoCPC is to make it available. I’m working with retailers in France, the UK, and Spain, and hope it’ll be on sale very soon.”

The post PicoCPC custom board appeared first on Raspberry Pi.

The pace of AI innovation continues to accelerate. Just a few months after the release of GPT-5.2 and GPT-5.3, Microsoft Foundry now brings us GPT-5.4—OpenAI’s most capable frontier model to date to Microsoft Foundry. This release marks a significant leap forward in reasoning, agentic workflows, and professional-grade automation.

Whether you’re building AI agents, automating complex workflows, or exploring new frontiers in software development and data analysis, GPT-5.4 is designed to deliver faster, more reliable, and more intelligent results.

1. What’s in GPT-5.4?
2. Introducing GPT-5.4 Pro
3. Capabilities
4. Pricing and Availability
5. Thoughts
6. Resources

What’s in GPT-5.4?

GPT-5.4 introduces a range of enhancements that elevate its performance across professional and enterprise scenarios:

• Built-in agentic workflows for planning and execution
• Native computer use capabilities (keyboard, mouse, screenshots)
• Tool Search for navigating large tool ecosystems
• Support for very large context windows (up to 1,050,000 tokens)
• Improved token efficiency for faster, lower-cost responses
• Enhanced coding and software automation reliability
• Higher factual accuracy and reduced hallucinations

These capabilities make GPT-5.4 a powerful tool for document and spreadsheet creation, coding, data analysis, and long-form reasoning tasks.

Introducing GPT-5.4 Pro

For scenarios where analytical depth and completeness are more important than latency, Microsoft Foundry also offers GPT-5.4 Pro—a premium variant of the model.

GPT-5.4 Pro is designed for deep analytical workflows, such as scientific research, strategic decision-making, and complex problem-solving. It introduces:

• Multi-path reasoning evaluation to explore alternative solutions
• Greater analytical depth for problems with trade-offs or multiple valid outcomes
• Improved stability across long reasoning chains
• Enhanced decision support where rigor outweighs speed

With a larger context window (400,000 tokens at the moment, larger context window coming soon) and the same high output capacity (128,000 tokens), GPT-5.4 Pro is the go-to model for high-assurance, high-complexity tasks.

Capabilities

To help you evaluate the evolution of these models, here’s a side-by-side comparison of GPT-5.4, GPT-5.4 Pro, and GPT-5.2:

Capability	GPT-5.4	GPT-5.4 Pro	GPT-5.2
Reasoning	Stronger reasoning for complex, multi-step tasks	Multi-path reasoning evaluation for deeper analysis	Adaptive reasoning for complex queries
Agentic Workflows	Built-in agentic workflows for planning and execution	Enhanced agentic workflows with improved stability	Accelerates agent development
Computer Interaction	Native computer use (keyboard, mouse, screenshots)	Same as GPT-5.4	Not available
Tool Management	Tool Search for large tool ecosystems	Same as GPT-5.4	Reliable tool use and governed integrations
Token Efficiency	Improved for faster, lower-cost responses	Same as GPT-5.4	Improved over GPT-5.1
Coding Reliability	Enhanced software automation and code generation	Same as GPT-5.4	Reliable code generation and modernization
Factual Accuracy	Higher factual accuracy and reduced hallucinations	Same as GPT-5.4	Greater consistency and accuracy
Context Window	1,050,000 tokens	400,000 tokens (1,050,000 coming soon)	400,000 tokens
Context Memory – Output	128,000 tokens	128,000 tokens	128,000 tokens
Training Data Cutoff	August 2025	August 2025	August 2025
Best For	Reliable execution, agentic workflows, software automation	Scientific research, complex decision-making, deep analytical workflows	Long-form reasoning, structured content, enterprise agents
Latency vs. Depth	Prioritizes speed and reliability	Prioritizes analytical depth and completeness over latency	Balanced performance

Pricing and Availability

Microsoft Foundry offers flexible pricing for GPT-5.4 based on context length:

• GPT‑5.4 (<272K input tokens):
  • $2.50 per million input tokens
  • $0.25 per million cached input tokens
  • $15.00 per million output tokens
• GPT‑5.4 (>272K input tokens):
  • $5.00 per million input tokens
  • $0.50 per million cached input tokens
  • $22.50 per million output tokens
• GPT‑5.4 Pro:
  • $30.00 per million input tokens
  • $180.00 per million output tokens

At launch, GPT-5.4 is available in Standard Global and Standard Data Zone (US), while GPT-5.4 Pro is available in Standard Global, with more deployment options expected soon.

Thoughts

GPT-5.4 is a great leap forward in enterprise AI. With its massive context window, built-in agentic capabilities, and native computer interaction, it’s built for the next generation of professional work. And for those who need even more analytical depth, GPT-5.4 Pro offers unmatched reasoning power.

The future of AI in the enterprise is arriving faster than ever—and Microsoft Foundry is making it easier to adopt and scale these capabilities securely and efficiently.

What are you most excited to build with GPT-5.4?

---

Resources

• Microsoft Tech Community Blog Post: https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-gpt-5-4-in-microsoft-foundry/4499785
• Microsoft Learn: https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure?tabs=global-standard-aoai%2Cglobal-standard&pivots=azure-openai#gpt-54

For more than three decades, Cindy Cohn, the executive director of the Electronic Frontier Foundation (EFF) has been at the center of the fight to protect privacy, free expression, and innovation online—taking on the NSA’s mass surveillance programs, defending encryption, and pushing back against efforts to weaken digital security in the name of safety. In her new book, Privacy's Defender, she reflects on the landmark cases that shaped the modern internet, the values that guide EFF’s work, and why privacy is not about hiding wrongdoing, but about preserving human autonomy and democracy in a networked world. Rainey Reitman, co-founder of the Freedom of the Press Foundation, leads our conversation.

Grab your copy of Privacy's Defender: https://mitpress.mit.edu/9780262051248/privacys-defender/ 

This conversation was recorded on 02/23/2026.

Check out all of the Future Knowledge episodes at https://archive.org/details/future-knowledge

Download audio: https://media.transistor.fm/fb8e3281/92372d70.mp3

Apple unveiled the MacBook Neo, the company's foray into a low-cost laptop. The iPhone Fold's supposed design has leaked through 3D CAD rendering files. And a toolkit for hacking iPhones has leaked.

• 18 years later, Apple ships a $599 computer.
• Apple's TikTok ads for the MacBook Neo are the right kind of weird.
• Apple creates adorable little Finder guy to promote its adorable little Mac.
• The new Apple begins to emerge.
• Apple 'Ultra' products expansion is up next after MacBook Neo launch.
• iPhone Fold design leaks in purported 3D CAD rendering files.
• Apple's 'HomePad' gets launch timing update via leaker.
• Apple's 512GB Mac Studio vanishes, a quiet acknowledgment of the RAM shortage.
• Apple Music to add Transparency Tags to distinguish AI music, says report.
• Apple ran a test on the App Store to see if AI could improve search result rankings.
• Apple geoblocking downloads of ByteDance-owned apps in the US.
• A toolkit for hacking iPhones, possibly created for the U.S. Government, has leaked.
• F1: The Stream - how the launch leveraged Apple's entire ecosystem.
• 'Apple' Review: Reinvention Incorporated.

Picks of the Week

• Christina's Pick: What's Your JND Game
• Andy's Pick: Kids, Wait Till You Hear This
• Jason's Pick: Cloth Pro Max
• Leo's Pick: Art Bits from HyperCard

Hosts: Leo Laporte, Andy Ihnatko, Jason Snell, and Christina Warren

Download or subscribe to MacBreak Weekly at https://twit.tv/shows/macbreak-weekly.

Join Club TWiT for Ad-Free Podcasts!
Support what you love and get ad-free audio and video feeds, a members-only Discord, and exclusive content. Join today: https://twit.tv/clubtwit

Sponsors:

• hipebl.ai
• helixsleep.com/macbreak

Download audio: https://pdst.fm/e/pscrb.fm/rss/p/mgln.ai/e/294/cdn.twit.tv/megaphone/mbw_1015/ARML9068113820.mp3

This week’s been wild — Iran bombed AWS data centers to take down Claude, OpenAI dropped GPT-5.4 (and it’s seriously good for coding), and living brain cells are literally playing DOOM. We’ve also got a heartfelt take on what it feels like to be a 10x engineer in the age of AI, plus some cool new tools like Handy for speech-to-text and web haptics. Oh, and new MacBook Pros with M5 Pro and M5 Max are up for pre-order. Try not to impulse buy (or do).

View the newsletter

Join the discussion

Changelog++ members support our work, get closer to the metal, and make the ads disappear. Join today!

Sponsors:

• Sonatype – Develop software fearlessly. Find out how at sonatype.com.

Featuring:

• Adam Stacoviak – Website, GitHub, LinkedIn, Mastodon, X

Download audio: https://op3.dev/e/https://pscrb.fm/rss/p/https://cdn.changelog.com/uploads/news/183/changelog-news-183.mp3