Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchAI, startups & growth insights. No spam.
The world's most powerful AI models are about to get a lot cheaper. Meta is moving toward releasing two frontier systems — internally codenamed Mango and Avocado — targeting a first-half 2026 window. Together they represent the most ambitious push yet into multimodal and code-generation AI. Whether they ship as open weights or behind an API paywall will determine which direction the industry tilts for years. The stakes are higher than any benchmark score: this is about who gets to build the next generation of software, and whether the tools to do it will be free.
To understand what Mango and Avocado mean, you need to understand what Llama 3 and Llama 4 already accomplished.
When Meta released Llama 1 in early 2023, it was a research artifact that leaked onto 4chan within days. That accident turned out to be one of the most consequential events in AI history. Engineers around the world who had never had access to a frontier-class language model suddenly did, and they immediately started fine-tuning it for everything — customer service bots, coding assistants, legal document review, local medical tools. The genie was out of the bottle.
Meta recognized the momentum and leaned into it. Llama 2 arrived as an officially open release. Llama 3 scaled up to 70 billion parameters and posted results that genuinely embarrassed closed-source competitors on several key benchmarks. Each generation expanded what was possible to build without an API credit card.
Then came Llama 4, released on April 5, 2025 — and it was a genuinely different kind of product. Two models shipped as open weights: Llama 4 Scout and Llama 4 Maverick. Both use a Mixture-of-Experts (MoE) architecture — activating only a fraction of their total parameters per inference pass, which dramatically reduces compute costs without sacrificing capability.
Llama 4 Scout offers 17 billion active parameters across 16 experts (109 billion total), fits on a single server-grade GPU, and ships with an industry-leading 10 million token context window on the instruction-tuned variant. Llama 4 Maverick runs the same 17 billion active parameters but spreads them across 128 experts (400 billion total), benchmarks favorably against GPT-4o and Gemini 2.0 Flash, and handles sophisticated multimodal inputs across text and images.
A third model — Llama 4 Behemoth — previews what training at true scale looks like: 288 billion active parameters, roughly 2 trillion total, with benchmark results that reportedly surpass GPT-4.5 and Claude Sonnet 3.7 on STEM tasks. Behemoth is still in training and not publicly available, but it exists — and it is the foundation on which Mango and Avocado were reportedly distilled.
The Llama 4 family landed on Hugging Face with 149 upvotes and 150+ community contributors within days. Fine-tuning forks proliferated. Infrastructure built around Llama 3 snapped into place for Llama 4 within weeks. This is what a mature open-source AI ecosystem looks like.
Mango is the more visually dramatic of the two upcoming models. Internally developed as Meta's next-generation image and video foundation system, Mango is designed to understand, generate, and edit visual media — not as a bolt-on capability but as its core function.
Based on reporting from The Information and TechCrunch, Mango is built on a multimodal diffusion-transformer architecture — the same class of model that underpins most of the leading video generation systems today, but reportedly with more sophisticated physical reasoning baked in. Early internal demos described in leaked accounts suggest the model maintains temporal coherence across 10-second sequences, a long-standing problem for video generation where objects routinely change shape, color, or mass between frames.
What makes Mango more interesting than a standard video generator is its reported emphasis on world modeling — the ability to reason about scenes, predict future frames, and allow users to manipulate variables like lighting or object placement while the model recalculates outcomes in real time. This is not text-to-video in the conventional sense. If the leaks are accurate, Mango operates closer to a physics-aware simulator than a creative media tool. That distinction matters enormously for enterprise applications: architectural visualization, product design, synthetic training data, and robotics simulation all need the kind of consistent physical grounding that current video generators fail to provide.
The deployment plan is equally ambitious. Mango is expected to power visual AI across Meta's entire platform — Facebook, Instagram, WhatsApp, and the Ray-Ban smart glasses line. When Meta deploys at Meta scale, "launch" is not a press release. It is a billion-user product update.
The open-source question for Mango is complicated. Meta's official statements have not confirmed whether Mango will follow the Llama tradition of open weights. Given that Mango is expected to become infrastructure for Meta's advertising and social products, there are business reasons to keep the weights proprietary. But there are also precedents in Meta's research history — including DALL-E competitor models and image segmentation work under SAM — where Meta released models publicly even when competitors did not.
Benchmark expectations are hard to set without confirmed specs, but the competitive field Mango enters includes OpenAI's Sora, Google DeepMind's Veo 3, and Runway's Gen-3. If Mango's physical reasoning claims hold up in independent testing, it would represent a meaningful leap over the current generation of video generators, most of which still struggle to keep a cup of coffee from teleporting across a table between frames.
Avocado is the model that has generated the most controversy inside and outside Meta — not because of what it does, but because of what it signals about who gets to use it.
Avocado is Meta's next-generation text and code model, being developed inside a unit called "TBD" within Meta's AI Superintelligence Labs. The group is led by Alexandr Wang, the 28-year-old founder of Scale AI, who became Meta's Chief AI Officer after Meta took a near-majority stake in Scale AI. Wang reportedly favors closed-source development, and Avocado is reportedly the first major model to reflect that preference.
On the technical side, Avocado targets the specific class of problems where current language models — including Llama 4 Maverick — still fall short: long-horizon software development, complex debugging, multi-file repository understanding, and production-quality code generation. Maverick's MBPP score of 77.6 and LiveCodeBench score of 43.4 are respectable but not dominant. The coding AI market — GitHub Copilot, Cursor, Windsurf, Devin — is one of the fastest-growing segments in enterprise software, and Meta has been conspicuously absent from it as a product.
Avocado is designed to change that. The model was reportedly trained using data drawn from multiple sources including outputs from Alibaba's Qwen, Google's Gemma, and other frontier systems — a distillation approach that has become increasingly common as the gap between model families narrows. The companion relationship between Mango and Avocado is also worth noting: the two models are reportedly designed to communicate through shared embeddings, enabling near-real-time prompt chaining where Avocado handles code and narrative scaffolding while Mango handles the visual execution. Together they function as a multimodal development environment, not just two separate models.
The timeline for Avocado has slipped. Multiple insiders told Bloomberg in December 2025 that the model faced internal delays related to capability gaps and organizational restructuring after Alexandr Wang's arrival. Spring 2026 remains the public target, but the development pace appears to be cautious rather than aggressive — which may reflect the scale of what's being attempted, or the difficulty of the pivot toward closed-source development within a company that has spent three years building an open-source identity.
For developers, Avocado's open-source status is existential. If Meta ships Avocado as a closed API, it competes directly with OpenAI and Anthropic for enterprise coding budgets. If it ships as open weights — even with usage restrictions — it instantly becomes the foundation for dozens of specialized coding tools that could undercut every commercial competitor.
The performance numbers matter less than the licensing terms. This has been true since Llama 1 leaked in 2023, and it remains true now.
When a model ships as open weights, the downstream effects are nonlinear. A research lab in Nairobi can fine-tune it for local legal precedent. A three-person startup in Lisbon can build a specialized coding assistant without an API budget. A hospital system in rural Indiana can run inference locally without sending patient data to a cloud provider's servers. None of these use cases require the model to be the absolute best in the world. They require it to be good enough, and free.
Llama 4 Scout and Maverick are both good enough for an enormous range of production applications. Their MoE architecture means they run efficiently on hardware that was prohibitively expensive just two years ago. The combination of open weights and efficient inference is what democratizes AI — not benchmark leaderboards.
This is why the Avocado situation is so consequential. If Meta's most capable coding and reasoning model stays behind a paywall, the open-source ecosystem loses its most important ally at the exact moment when the gap between open and closed models is narrowing fastest. The community has already demonstrated it can close performance gaps through fine-tuning, quantization, and synthetic data generation. What it cannot manufacture is compute — and closed models deny access to the weights that make efficient fine-tuning possible.
The counterargument is that Meta has every incentive to maintain developer goodwill, and that goodwill comes from open releases. Llama's reputation has driven AI Platform adoption, attracted research talent, and generated more press than any marketing campaign. Abandoning that for short-term API revenue is a real strategic gamble — especially while Meta navigates the broader hardware landscape and competes on inference cost.
Zuckerberg has been unusually candid about Meta's open-source rationale: every dollar companies spend on AI model access via API is a dollar not spent on Meta's advertising products. Cheap, capable AI infrastructure is effectively a tax on every competitor that builds AI products, because those competitors pay OpenAI or Anthropic for frontier access. Open-sourcing removes that lever from rivals.
This calculus gets more complicated at the frontier. Zuckerberg has shifted his public position on superintelligent models, citing safety concerns he never raised when defending Llama 2 or Llama 3. The $70-72 billion capital expenditure Meta committed for 2025 infrastructure — data centers, custom silicon, the Prometheus Data Center — is not being built to host an API product. It is being built to train models at scale, then selectively release what serves Meta's platform goals while potentially withholding capabilities that represent genuine competitive advantage.
Mango, as infrastructure for Instagram Reels and Facebook's creative tools, may never need open-sourcing because its value to Meta is in deployment. Avocado, as a potential GitHub Copilot Enterprise competitor, may stay closed to justify enterprise pricing. The business logic is coherent. Whether it serves developers is a different question.
The H1 2026 window is roughly three to six months from the writing of this article, which means both Mango and Avocado could ship before the summer. Based on the available reporting, here is the most credible sequence:
Q1 2026 (now through March): Internal capability evaluations, safety red-teaming, and final pre-release testing. Both models are reportedly in this phase. No public beta access is expected during this window based on current signals.
April-May 2026: The most likely release window for at least one of the two models, possibly Avocado first given its code-focused nature and the enterprise demand for capable coding AI. Meta tends to time major AI releases to developer conferences and the spring news cycle.
May-June 2026: Mango, which requires more infrastructure for video inference at scale, likely follows Avocado by weeks or months. Platform integration across Facebook, Instagram, and WhatsApp adds deployment complexity that pure language model releases do not face.
Developers who want to prepare should be running Llama 4 Scout and Maverick locally now — the MoE architecture, the multimodal input pipeline, and the tokenization approach will likely carry forward into Avocado if it follows the Llama family architecture. Building applications on Llama 4's API surface today means less migration work when Avocado arrives.
The Hugging Face model hub is the canonical source for release announcements and weight downloads when they come.
The Mango and Avocado announcements — even at the level of rumor and internal leak — have reshaped the competitive calculus in ways that are already visible.
OpenAI's accelerated GPT-5.4 release came with unusually aggressive pricing, suggesting the company is anticipating margin compression from open-source competition. When open weights cover 80 percent of what a commercial API does at zero cost, closed-source providers can only defend on capability gap or ecosystem lock-in. OpenAI is pursuing both: scaling for capability, Operator integrations and ChatGPT distribution for lock-in.
Anthropic's position is more nuanced. Claude's differentiation has always been safety-focused reasoning and enterprise trust, not raw benchmark performance. If Avocado ships as an open-weight coding model, it threatens GitHub Copilot and Cursor more directly than it threatens Claude Enterprise. Anthropic's bet is that enterprises willing to pay for AI will pay for reliable AI with clear liability frameworks — not just the cheapest capable option.
Google's response has been the most aggressive on the infrastructure side. The TPU deal and NVIDIA competition dynamics that have been playing out in 2026 reflect Google's awareness that the model quality gap is narrowing, and that the next competitive dimension is inference cost and hardware access. Gemini 2.0 Flash and the Gemma open-weight family are Google's answer to Llama — capable models released freely to build developer goodwill while the company monetizes inference through Google Cloud.
DeepSeek's open-source trillion-parameter work complicates the picture further. When a Chinese lab can release frontier open weights at a fraction of US development costs, every closed-source position faces pressure. Meta's open-source strategy is partly a response: if frontier open weights will exist regardless, Meta would rather be the one releasing them than the one being undercut.
The open-source community around Llama is more capable than ever. On Hugging Face, the Llama 4 Scout and Maverick releases attracted 150+ community contributors within days. Transformers v4.51.0 added full MoE architecture support. TGI shipped optimized kernels for Maverick's 128-expert routing within a week. Quantization tooling for int4 and FP8 inference is mature and production-tested.
Hardware costs have dropped sharply. Llama 4 Scout runs on a single H100 GPU in 4-bit quantization at roughly $3/hour on major cloud providers. For larger successors like Avocado, distributed inference frameworks — vLLM, LiteLLM — are already hardened on Llama 4. Multi-GPU serving tooling exists for Behemoth-scale models.
Fine-tuning will follow the established Llama playbook: LoRA and QLoRA adapters, synthetic data generation, domain-specific benchmarks. Avocado fine-tuned on a company's internal codebase with retrieval-augmented generation will likely outperform any generic commercial coding API for that organization's specific use cases — assuming the weights are available to fine-tune.
The honest answer is: they already are, regardless of what Meta decides about Mango and Avocado's licensing.
The Llama 4 generation crossed a threshold previous open-weight releases had not. Scout and Maverick are not "good for open source" — they are simply good. They beat commercial models on multiple benchmarks, run efficiently on accessible hardware, support genuine multimodal workflows, and ship with a 10 million token context window that no commercial API matched at comparable price. The usual caveat — "impressive for an open model" — no longer applies.
What Mango and Avocado represent is the question of whether that progress continues into the next capability tier. If both ship as open weights, the ecosystem gains access to frontier-class multimodal and coding systems that currently cost hundreds of dollars per month via API. The downstream innovation would be staggering: startups building on open foundations, researchers fine-tuning for specialized domains, enterprises deploying without data leaving their infrastructure.
If either model ships closed, the community will adapt. DeepSeek has demonstrated that open-source AI development is not dependent on any single organization. But Meta's infrastructure advantages — the training compute, the hardware investment, the 10-year head start in social media data — are real. A closed Meta frontier model is a meaningful setback for open-source AI in a way that a closed Anthropic or OpenAI model is not, because Meta was the one company large enough to sustain open-source development at the frontier.
The decision Meta makes about Avocado's licensing will be one of the most consequential choices in AI's near-term history. It will not end open-source AI — that ship sailed when Llama 1 leaked in 2023. But it will determine how fast the open ecosystem catches the closed one, and how many builders worldwide have access to the tools that define what software looks like in 2027.
Open source has already won the argument. The question now is whether it wins the race.
For more on Meta's hardware strategy context, see Meta and Google's TPU deal and what it means for NVIDIA. For how this fits into the broader open-source model landscape, see DeepSeek V4's trillion-parameter open-source push. And for how the competition is responding on the closed-source side, see OpenAI's GPT-5.4 launch.
Sources: Meta AI Blog · Hugging Face Llama 4 Release · TechCrunch – Meta Llama 4 · TechCrunch – Mango · Engadget – Avocado
Meta's first Llama developer conference unveils a hosted Llama API, 1 billion downloads milestone, Llama 4 Scout and Maverick, and a new SDK that rivals OpenAI's platform.
MiniMax released M2.5 — a 230 billion parameter mixture-of-experts model with 80.2% on SWE-Bench Verified, matching Claude Opus 4.6 speed at $0.29/M input tokens, trained across 200,000 real-world coding environments.
Indian AI lab Sarvam AI releases open-source 30B and 105B parameter LLMs with 22-language support, speech, and vision — challenging closed-source AI dominance.