Everything we announced at our first LlamaCon

Original Article ↗ Hacker News Discussion ↗

Timing, Qwen 3, and Missing “Thinking” Model

  • Several commenters speculate Meta held back a Llama 4 “Thinking”/reasoning model because Qwen 3 launched the same day with strong benchmarks and many local variants.
  • Some see this as Qwen “bullying” Meta with timing; others say Meta’s “unlucky timing” isn’t just luck.
  • There is notable skepticism about Qwen 3’s advertised benchmarks (particularly small models outperforming prior large ones), with suggestions that distillation/RL may be leaking benchmark data into models.

Meta’s Strategy: Cloud, API, and Platform Play

  • The Llama API preview is seen as Meta moving closer to cloud/platform economics rather than just social networks.
  • Some think this is Meta trying to own more of the AI stack, as Llama is already first-class on major clouds and they may want a bigger share than licensing alone.
  • Others find the announcement list underwhelming: lots of positioning around speed and integrations but few concrete numbers or new models; security tools (Llama Guard 4, LlamaFirewall, Prompt Guard 2) are considered the most substantive.
  • SAM 3 (Segment Anything 3) is noted as an upcoming highlight outside the core Llama API story.

Perceptions of Llama Models and Real-World Use

  • Mixed views on Llama’s competitiveness: some call it “subpar” and note Llama 4’s weakness at coding vs models like DeepSeek R1 and newer Chinese models.
  • Others report successful use of Llama 3.3 locally for tasks like large-scale document labeling, claiming better results than Phi-4, Gemma 3, or Mistral Small under constrained hardware.
  • In local-LLM communities, several say the momentum has shifted toward Gemma, Qwen, and Mistral; others caution not to over-read “hive mind” sentiment from Reddit/Discord and advocate own-stack testing.

Specialized vs General LLMs

  • A question about domain-specific Llama variants (e.g., pure programming models) leads to the view that:
    • Specialized models exist (e.g., code/math), but general models plus Mixture-of-Experts tend to catch up and outperform niche models within months.
    • Non-technical data still helps with reasoning, and multi-domain/multi-language training appears to improve overall capability.
    • Fine-tuning on custom datasets is framed as the practical route for specialization.

Open Source, Licensing, and “Open-Washing”

  • Strong criticism of Meta’s repeated “open source” framing: commenters note the Llama license’s commercial restrictions and advertising clause conflict with the Open Source Definition.
  • Multiple people argue that without training data and with license constraints, Llama is at best “open-weights” or “weight-available,” not open source.
  • Some point out many other big tech firms release less restricted open-weight models, questioning Meta’s positioning as uniquely “open.”
  • There’s debate over whether model weights are “binary blobs” and how copyright applies, with a long subthread on whether licensing weights is even meaningful in traditional IP terms.
  • A few suggest Meta is in a bind: serious license enforcement could trigger a community backlash and migration to truly open models.

Trust, Privacy, and Meta’s Broader Ambitions

  • One thread imagines Meta using AI as a wedge into smart homes: local, privacy-preserving assistants tied into social/commerce.
  • This vision meets heavy skepticism: several argue Meta is poorly suited to “building privacy and trust,” citing its ad/surveillance-driven business model and history of manipulation.
  • Others counter that in regions like Southeast Asia, Meta’s services are deeply embedded in everyday life and commerce, functioning as de facto marketplaces in the absence of Amazon/Shopify-style infrastructure.
  • A subset warn that any “free” AI from Meta will ultimately be tied to harvesting and monetizing user “memory” and behavior.

Metaverse and Hardware Side Discussion

  • Some express surprise that AI efforts survived the metaverse pivot and interpret “all AI, all the time” as Meta quietly de-emphasizing VR world-building.
  • Others note ongoing XR efforts (Quest 3, rumored HUD glasses) and describe real productivity use cases with VR desktops.
  • Comparisons with Apple Vision Pro surface: better desktop UX vs Quest for some, but comfort, balance, and weight are recurring pain points.

Miscellaneous

  • A few users report practical issues signing up for the new Llama API waitlist (login redirect loops).
  • Some disappointment that no multimodal/“Omni”-style, voice-native Llama was announced.