Up to date
meta-llama Llama 3.2 Community License out 2 years ago
3.21B parameters, text in, text out, 128K tokens of context. 3 providers serve it from $0.050 per million input tokens.
Verified 2026-10-25 Rank #16 in LLMs · score 120.2 method v1 Sources 5 Downloads 30d 1.1M · #103 of 619 tracked
Changed this week: 2 new quantized builds on the Hub (542 in all).2026-10-24
No drawable SVG in the reply.
Why did the open-weights model cross the road?
To get to the other side, but it was a little "light" on its feet.
___________ | | | _______ | | / \ | |/________\| | | | _______ | | | | | | | PARASAILER | | | | | | |_______| | | | |___________|
No playable frames in the reply.
The ringer: the same four playful tests for every model, replies shown as received and checked by code. A joke, not a benchmark; never ranked.0 of 4 pass · parasail-llama-32-3b-instruct on parasail · run 2026-10-09 · temperature 0 · cost $0.011 · 2026-10 thing: supermarket · all models →
Picked by rule from 34 tracked pieces: the lab's own post, the most shared, a field report, the latest.
The project trained a Sparse Autoencoder on 6.5 Billion activations from Llama 3.2-3B layer 23 over 10 epochs.
A guide details self-hosting Llama 3.2 on a home server using Coolify and Ollama.
The 3 picks above, plus these 22. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.
Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.
2 providers with a live price, from $0.050 to $0.051 per million input tokens; the median is $0.050. One dot per provider, cheapest in teal.
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
Served by 3 API providers; heavily run locally (542 quantized builds); a big base for builders (2,939 fine-tunes, adapters and merges).
Top Spaces running it: pliny-the-prompter/obliteratus ♥ 621 · cvachet/pdf-chatbot ♥ 382 · jupyter-agent/jupyter-agent ♥ 321 · baconnier/prompt-plus-plus ♥ 246
Weights only, an estimate: leave room for the KV cache. Method
About 1.9 GB at INT4, 3.9 GB at FP8, 7.7 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
Third-party measurements only; author-reported numbers stay in Reference, marked as such.
21.95. Each tick is a model on this site with a measured BFCL row; author-reported numbers stay in Reference. Percent (overall accuracy). FC
Source · leaderboard · 2026-10-25
Claim-matched tests, graded by code. Receipts, not a score.
{
"providers": [
{
"provider": "parasail",
"endpoint_model": "parasail-llama-32-3b-instruct",
"source_url": "https://api.parasail.io/v1/chat/completions",
"ttft_ms_median": 456,
"ttft_any_ms_median": 456,
"tokens_per_s_median": 200.8,
"runs": [
{
"ttft_s": 0.66,
"ttft_any_s": 0.66,
"tokens_per_s": 159.9,
"latency_s": 1.15,
"tokens_out": 184,
"reasoning_chars": 0,
"finish": "stop"
},
{
"ttft_s": 0.441,
"ttft_any_s": 0.441,
"tokens_per_s": 200.9,
"latency_s": 0.92,
"tokens_out": 184,
"reasoning_chars": 0,
"finish": "stop"
},
{
"ttft_s": 0.456,
"ttft_any_s": 0.456,
"tokens_per_s": 200.8,
"latency_s": 0.92,
"tokens_out": 184,
"reasoning_chars": 0,
"finish": "stop"
}
]
}
],
"note": "Single client, wall clock; tok/s counts every output token including hidden reasoning."
} Weights, every dated article, and repo commits on one rail.
Teal is the weights (2024-09-18), amber triangles are 8 dated articles (each links to it).
3,481 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Everything else, folded: the full tables, every source.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| measured BFCL · Call tools and APIs | 21.95 percent (overall accuracy) | FC5 sub-scores
| leaderboard | 2026-10-25 |
Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.
Listed in LLMs
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | LlamaForCausalLM |
| Model type | llama |
| Params by dtype | BF16: 3,212,749,824 |
| Size band | M |
| MoE | no |
| Context length | 131,072 (provider listing) |
| Gated | manual |
| Papers | arXiv:2204.05149, arXiv:2405.16406 |
| License | Llama 3.2 Community License (llama3.2) |
| Hub snapshot | 2026-10-25 |
vLLM: supported · LlamaForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · LlamaForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed
vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.
Weights-only memory for Llama-3.2-3B-Instruct: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 7.7 GB · Single GPU (<24 GB) |
|---|---|
| FP8 (1 byte/param) | 3.9 GB · Edge (<4 GB) |
| INT4 (0.5 bytes/param) | 1.9 GB · Edge (<4 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.
quantized: 542finetune: 2,038merge: 29adapter: 872
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF | quantized | GGUF | — | — | — | MaziyarPanahi | 15 | 10M |
| bartowski/Llama-3.2-3B-Instruct-GGUF | quantized | GGUF | — | — | llama3.2 | bartowski | 251 | 4.9M |
| unsloth/Llama-3.2-3B-Instruct | finetune | — | 3.21B | M | llama3.2 | unsloth | 99 | 4.3M |
| inference-net/Schematron-3B | finetune | — | 3.21B | M | llama3.2 | inference-net | 345 | 3.5M |
| unsloth/Llama-3.2-3B-Instruct-unsloth-bnb-4bit | quantized | -bnb | 3.30B | M | llama3.2 | unsloth | 11 | 2.4M |
| unsloth/Llama-3.2-3B-Instruct-bnb-4bit | quantized | -bnb | 3.30B | M | llama3.2 | unsloth | 37 | 2.1M |
| unsloth/Llama-3.2-3B-Instruct-GGUF | quantized | GGUF | — | — | llama3.2 | unsloth | 89 | 1.1M |
| mlx-community/Llama-3.2-3B-Instruct | finetune | — | 3.21B | M | llama3.2 | mlx-community | 10 | 998.6k |
| RedHatAI/Llama-3.2-3B-Instruct-FP8-dynamic | quantized | -FP8 | 3.61B | M | llama3.2 | RedHatAI | 3 | 941.7k |
| lmstudio-community/Llama-3.2-3B-Instruct-GGUF | quantized | GGUF | — | — | llama3.2 | lmstudio-community | 48 | 835.8k |
| RedHatAI/Llama-3.2-3B-Instruct-FP8 | quantized | -FP8 | 3.61B | M | llama3.2 | RedHatAI | 6 | 532.1k |
| hugging-quants/Llama-3.2-3B-Instruct-Q4_K_M-GGUF | quantized | GGUF | — | — | — | hugging-quants | 49 | 472.8k |
| hugging-quants/Llama-3.2-3B-Instruct-Q8_0-GGUF | quantized | Q8_0 | — | — | — | hugging-quants | 53 | 306.3k |
| SanctumAI/Llama-3.2-3B-Instruct-GGUF | quantized | GGUF | — | — | llama3.2 | SanctumAI | 2 | 233.5k |
| X-iZhang/libra-v1.0-3b | merge | — | 3.25B | M | apache-2.0 | X-iZhang | 2 | 196.2k |
| unsloth/Llama-3.2-3B-unsloth-bnb-4bit | quantized | -bnb | 3.30B | M | llama3.2 | unsloth | 4 | 195.4k |
| canopylabs/orpheus-3b-0.1-pretrained | finetune | — | 3.78B | M | apache-2.0 | canopylabs | 175 | 167k |
| mlc-ai/Llama-3.2-3B-Instruct-q4f32_1-MLC | quantized | — | — | — | — | mlc-ai | 2 | 142.8k |
| acon96/Home-Llama-3.2-3B | finetune | GGUF | 3.21B | M | other | acon96 | 18 | 136.5k |
| QuantFactory/Llama-3.2-3B-Instruct-abliterated-GGUF | quantized | GGUF | — | — | llama3.2 | QuantFactory | 36 | 125.5k |
| mlc-ai/Llama-3.2-3B-Instruct-q4f16_0-MLC | quantized | — | — | — | — | mlc-ai | 2 | 120.9k |
| mlc-ai/Llama-3.2-3B-Instruct-q4f16_1-MLC | quantized | — | — | — | — | mlc-ai | 5 | 101.5k |
| second-state/Llama-3.2-3B-Instruct-GGUF | quantized | GGUF | — | — | llama3.2 | second-state | 4 | 78.4k |
| huihui-ai/Llama-3.2-3B-Instruct-abliterated | finetune | — | 3.61B | M | llama3.2 | huihui-ai | 104 | 73k |
| radheneev/NetrAI-L3 | finetune | — | 3.21B | M | llama3 | radheneev | 3 | 71.7k |
| RedHatAI/Llama-3.2-3B-Instruct-quantized.w8a8 | quantized | -quant | 3.61B | M | llama3.2 | RedHatAI | 1 | 68.8k |
| uzlm/alloma-3B-Instruct | finetune | — | 3.21B | M | llama3.2 | uzlm | 6 | 68k |
| HKUSTAudio/Llasa-3B | finetune | — | 4.01B | M | cc-by-nc-4.0 | HKUSTAudio | 531 | 37k |
| QuantFactory/Llama-Sentient-3.2-3B-Instruct-GGUF | quantized | GGUF | — | — | creativeml-openrail-m | QuantFactory | 2 | 49.6k |
| QuantFactory/VersatiLlama-Llama-3.2-3B-Instruct-Abliterated-GGUF | quantized | GGUF | — | — | cc-by-4.0 | QuantFactory | 6 | 49.2k |
| QuantFactory/Llama3.2-3B-Enigma-GGUF | quantized | GGUF | — | — | llama3.2 | QuantFactory | 4 | 41.1k |
| QuantFactory/Llama3.2-3B-Esper2-GGUF | quantized | GGUF | — | — | llama3.2 | QuantFactory | 3 | 38.6k |
| espressor/meta-llama.Llama-3.2-3B-Instruct_W4A16 | quantized | — | 3.63B | M | — | espressor | 0 | 37.9k |
| canopylabs/3b-de-pretrain-research_release | finetune | — | 3.78B | M | llama3.2 | canopylabs | 14 | 29.9k |
| Skywork/Skywork-Reward-V2-Llama-3.2-3B | finetune | — | 3.21B | M | llama3.2 | Skywork | 5 | 28.3k |
| QuantFactory/Llama-3.2-Taiwan-Legal-3B-Instruct-GGUF | quantized | GGUF | — | — | llama3.2 | QuantFactory | 11 | 25.1k |
| legraphista/Llama-3.2-3B-Instruct-IMat-GGUF | quantized | GGUF | — | — | llama3.2 | legraphista | 2 | 25.3k |
| ValiantLabs/Llama3.2-3B-ShiningValiant2 | finetune | — | 3.21B | M | llama3.2 | ValiantLabs | 6 | 25.1k |
| CarrotAI/Llama-3.2-Rabbit-Ko-3B-Instruct | finetune | — | 3.21B | M | llama3.2 | CarrotAI | 11 | 19.8k |
| suayptalha/DeepSeek-R1-Distill-Llama-3B | finetune | — | 3.21B | M | llama3.2 | suayptalha | 16 | 19.1k |
Hosted endpoints observed for Llama-3.2-3B-Instruct, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.05/M.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| Parasail parasail, openrouter | parasail-llama-32-3b-instruct | 128K | bf16 | $0.05 | $0.33 | — | — | — | serverless | 2026-10-24 |
| Cloudflare openrouter | meta-llama/llama-3.2-3b-instruct | 78K | unknown | $0.05 | $0.34 | — | — | — | live | 2026-10-24 |
| Featherless hf | meta-llama/Llama-3.2-3B-Instruct | — | — | — | — | — | — | — | live | 2026-10-25 |
Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:31 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 14:59 UTC · Last verified 2026-10-25
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.