Meta Llama 3

Original Article ↗ Hacker News Discussion ↗

Model capabilities & benchmarks

  • Llama 3 released in 8B and 70B parameter variants; both are reported as major jumps over Llama 2.
  • 8B reportedly beats Llama 2 70B and Mistral 7B on many instruction benchmarks; 70B is said to outperform Claude 3 Sonnet, Gemini Pro 1.5, Mixtral 8x22B, and be competitive with some closed “mid-tier” models.
  • Shared GPT‑4(-Turbo) numbers show 70B still below GPT‑4 on reasoning-heavy benchmarks, but close enough that many expect strong practical performance, especially after community fine‑tuning.
  • A still‑training 400B dense model checkpoint already approaches GPT‑4 scores on several public benchmarks, and in some comparisons slightly beats original published GPT‑4 numbers.

Context window & use cases

  • Current context is 8k tokens. Some see this as inadequate versus 200k+ competitors for RAG, long documents, or complex formatting, though Meta says longer windows are coming.
  • Others argue very long contexts are overrated, costly, and slow prefill; many real tasks work fine with 8k and chunking.

Open weights vs “open source”

  • Strong debate: Meta calls Llama “open,” but several commenters stress the license is not OSI‑compliant (usage restrictions, at‑scale clauses, AUP), and that weights alone are more like “weights‑available” or “corporate freeware” than true open source.
  • Counter‑view: for almost all startups and individuals (under 700M MAU), the license is effectively permissive and practically as useful as open source.
  • Deeper arguments over whether reproducibility, training data, and training code are required for an “open” model; examples of more fully open projects are cited.

Availability, regulation & geofencing

  • Meta’s web assistant is accessible without login in some countries, but many users in EU, UK and others get “not available in your country.”
  • Explanations vary: EU privacy/AI regulation workload, legal risk, or simple phased rollout. Some note other vendors (Claude, Gemini) similarly exclude much of Europe.
  • There is disagreement over whether EU rules are “draconian” or simply force privacy‑exploiting firms to “actually care.”

Ecosystem impact & competition

  • Many see Llama 3 as a huge win for the open‑weight ecosystem: better local models, cheaper startups, faster experimentation, and erosion of proprietary moats.
  • Some speculate on winners/losers: possible pressure on OpenAI/Google, ambiguous effect on Nvidia (less training vs more fine‑tuning and inference), potential upside for others selling inference hardware.
  • Meta’s strategy is widely read as commoditizing foundation models to strengthen its core ad/social products and weaken closed competitors, not pure altruism.

Running locally & hardware

  • 8B runs comfortably on consumer GPUs (e.g., ~8–16GB with quantization; even laptops and Macs via tools like llama.cpp/ollama).
  • 70B can be run with heavy quantization on high‑end consumer GPUs or large‑RAM CPUs, but with modest token/s throughput; cloud A100/H100s give much better performance.
  • For the future 400B model, commenters expect it to be practically server‑only, though aggressive quantization and CPU inference may make slow local use possible.

Meta.ai product behavior & UX

  • Web console is praised for speed and relatively light‑handed censorship compared to some rivals; it can use live web search (Bing/Google) and shows citations.
  • Image generation (/imagine) updates in near‑real‑time as you type; the preview uses a different, faster model and may differ in style from the final images.
  • Some users report odd behaviors:
    • Long, repetitive “goodbye” loops and tokenization glitches in early 8B builds.
    • Generated non‑English answers being immediately deleted and replaced by “I don’t understand [language] yet” messages, likely due to server‑side filters.
    • Occasional self‑hallucinations about which model it is and what it can accept (e.g., PDFs, languages).

Licensing clauses & branding

  • New license requires redistributors and downstream models to:
    • Include the license text;
    • Prominently display “Built with Meta Llama 3”;
    • Prefix derivative model names with “Llama 3”.
  • Some see this as mild and comparable to open‑source attribution clauses; others say it’s a significant barrier for certain commercial or white‑label use.

Sentiment toward Meta & “openness”

  • Many are surprisingly positive about Meta’s role: long track record of major open projects, willingness to ship strong open‑weight LLMs, and resistance to “AI doomerism” used as a justification for complete closedness.
  • Others remain deeply skeptical given Meta’s history with privacy, engagement‑driven harms, and note that Meta has already said they won’t release weights once models are “too powerful,” expecting a future pivot to an OpenAI‑like posture.