Llama
No written verdict yet. Here is what commenters said.
Why people recommend it
“Did you forget what sparked the whole open-weight wave? It was that.”
“spin up a base (not instruction tuned) checkpoint of Llama 1”
Show 3 more
“Llama 3 8B on a chip”
“I've actually seen better performance with older ones like Meta Llama 4 or Amazon's Nova”
“I've gotten decent results with Llama 3.1 8b on Hugging Face tester with Exa MCP”
Caveats and dissent
“LLama 3 405B had the most unoptimized kv cache usage by far.”
“Except that Llama's license has restrictions that make it incompatible with the Open Source Definition.”
Show 2 more
“Llama has restrictions incompatible with the definition of Open Source.”
“Big turnaround from Llama 4.”
By version sentiment of people who named that release, oldest release first
Versions with at least 5 people. 28 people didn't name a version. Small samples swing a lot, so read the counts too. Compare all models →
Often compared with
Who recommends it in their own words
Threads where it came up
- Title: Show HN: Aurora – Live dashboard watching local LLMs create autonomously 3 mentions
- Mark Zuckerberg attacks 'closed' AI rivals as Meta returns to open models 2 mentions
- AMD acquires Taalas to boost inference performance by etching models in silicon 2 mentions
- Natural Language Autoencoders: Turning Claude's Thoughts into Text 2 mentions
- Granite 4.1: IBM's 8B Model Matching 32B MoE 2 mentions
- RWKV-7 beats Llama 3.2 with 3x fewer training tokens and formally exceeds TC^0 2 mentions
Also written as Llama 3.1 8B, Llama 3, Llama 3.3 70B, Llama 3.1. Data: llama.json. Quotes are verbatim; the verdict and theme labels are written by a model from these comments, and every clause links to its sources.
All comments 58
“I've tested pairs (like Llama 3 base vs. instruct)”
“pushing me towards the Abliterated (uncensored) models Llama-3-70B-Abliterated or Hermes 3 70B, Dolphin, R1-Abliterated”
“Llama 3.1 8B baked into an ASIC”
“Did you forget what sparked the whole open-weight wave? It was that.”
“spin up a base (not instruction tuned) checkpoint of Llama 1”
“Sure, it's just a 8B model or so (Llama 3.1 8B)”
“gemini-2.5-pro, claude-opus-4.8, gpt-5.1, llama-3.3-70b”
“llama 3.1 8b q8_0 does about 70tk/s”
“https://chatjimmy.ai/ runs Llama 3.1-8B on an ASIC”
“I have been using open source LLMs locally for the past 5 months like Qwen, Llama, DeepSeek and others”
“LLama 3 405B had the most unoptimized kv cache usage by far.”
“Except that Llama's license has restrictions that make it incompatible with the Open Source Definition.”
“Llama has restrictions incompatible with the definition of Open Source.”
“Qwen models alone have overtaken the previously-dominant Llama models in hf downloads by quite a margin.”
“Llama 3 8B on a chip”
“It's a quantised 8B model (Llama 3.1 8B to be exact).”
“Some small 8b models (like llama3.1) for sentiment analysis on text.”
“It's quite an old model now and small”
“I've actually seen better performance with older ones like Meta Llama 4 or Amazon's Nova”
“I've gotten decent results with Llama 3.1 8b on Hugging Face tester with Exa MCP”
“it's running the rather outdated Llama 3.1 8B”
“With some outliers like Mistral and Llama on some questions.”
“Big turnaround from Llama 4.”
“The Llama family just aren't very useful.”
“My project uses 2 bit quantization to run llama3 models on any device with enough ram.”
“It's not the better model since Llama3.”
“the Llama and Qwen models gained a lot”
“llama was unintentionally released and has been a flash in the pan for them since”
“Llama 3.3 Instruct 70B is 5-10x cheaper than Sonnet 5.”
“llama is not anywhere near the Pareto distribution (performance vs cost)”
“So practically dead!”
“Engaging in dialogues with three 2024 foundation AI models—Llama 3.1 405B, ChatGPT-4o, and Claude 3.5 Sonnet (new)”
“Why didn't this author compare Llama 3 with GLM 5.2”
“(1 response, from Llama 3.2 1B, classed as a hallucination)”
“it's llama 3.1 8B at 16000 tokens per second”
“transformed Llama 7B into a credible facsimle of ChatGPT”
“The underlying model is only llama 3 8B”
“Currently they're using a LLama 8B model.”
“reaches over 14k TPS on Llama 3.1 8B”
“Llama 4, and most notably the Grok 3 drama”
“Qwen 3.6 35b a3b, granite4.1 8b, llama 3.2 3b.”
“llama 3.2 3b is pretty decent for not hogging resources”
“For microscopic models I like the llama 3.2 3b more right now”
“Llama 4 was a bad model, unfortunately.”
“a LangGraph + Groq Llama 3.3 70B agent that reads GitHub PRs and issues”
“get Llama 3.2 1B running at 40-46 tok/s at 2W”
“Meta confirmed Llama 4 benchmarks were fudged.”
“a local True RAG demo using Llama 3.2”
“I sent 89 identical coding prompts to 6 LLMs”
“Llama 3.3 70B's higher success rate comes down to it reliably calling the details endpoint first”
> Specifically, we collected new data created after January 2025, including: [...] new fiction on Archive of Our Own (Various, 2025), Not sure how to feel about this. From a researcher's point of view, reproducibility is important, but the last time someone publicly collected data from AO3, the community was not very fond of that. https://huggingface.co/datasets/nyuuzyou/archiveofourown/discussions
“vs LLaMA 3.2's 69.7%”
“I have run this thousands times against Llama-4-Maverick-17B-128E-Instruct-FP8 and Gemini-3-Flash with consistently working results.”
“You can run smaller models locally that are pretty good at code gen.”
“serving LLama 3.3 70B”
“The same thing was done with Meta researchers with Llama 4”
Shipping GPU-accelerated ML models in games looks difficult, are there any major examples other than vendor-locked upscaling like DLSS or FSR? (Yep! https://cod.ifies.com/voxel-visibility/ )
“The "big" one is Llama3.3-70b on the cloud, right now.”