Model Follower.

Up to date

Rebuilt
2026-10-26 07:36 UTC
Release radar
116 on watch
Hub snapshot
2026-10-26
Re-checked
daily · every number sourced and dated

Models › Laguna family

Laguna-XS-2.1

Hugging Face ↗ Follow changes

poolside out 4 months ago

33.4B parameters (8 of 256 experts per token), text in, text out.

Verified 2026-10-26 07:27 UTC Rank #690 in LLMs · score 72.4 method v1 Sources 4 Downloads 30d 577.2k · #153 of 619 tracked

Can I run it?20GB at INT4Estimatefits a 24 GB cardmethod
Do people use it?577.2kdownloads, 30 days#153 of 619 we track · top 25%260 likesHugging Face

Worth reading

Picked by rule from 12 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 416 HN pts

Laguna S 2.1

Poolside releases Laguna S 2.1, an agentic coding model with 118B total parameters and 8B activated parameters per token that launched in under nine weeks.

  • Laguna S 2.1 scored 40.4% in our agent harness, pool, not mini-swe-agent which DeepSWE's leaderboard uses.
poolside.ai 2026-07-21 416 HN pts · 88 comments 3.8k words about the Laguna family family

12 best of 12 tracked

The 3 picks above, plus these 9. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleLaguna XS.2 and M.1poolside.ai · 2026-04-28 · 101 HN pts
  2. DiscussionAnybody tried Laguna S 2.1 (by Poolside)?news.ycombinator.com · 2026-08-06 · 7 HN pts
  3. ArticleAgentic AI at Two Different Scales: Nanbeige4.2-3B and Laguna S2.1kaitchup.substack.com · 2026-07-31 · 2 HN pts
  4. ArticleLaguna XS.2 and Laguna M.1 by Poolsidepoolside.ai · 2026-04-28 · 5 HN pts
  5. ArticleLaguna M.1/XS.2 Technical Report [pdf]poolside.ai · 2026-05-26 · 2 HN pts
  6. ArticleWatch Tesla's Record-Breaking Laguna Seca Lapengadget.com · 2019-09-12 · 6 HN pts
  7. ArticleLaguna XS.2: A Coding Model Built for Closed Environments, Now Open Weightsfirethering.com · 2026-04-29 · 1 HN pts
  8. VideoTesla Model S at Refuel 2013 (Laguna Seca)youtube.com · 2013-09-19
  9. DiscussionTesla is attempting lap record with a Model S Plaid+ at Laguna Seca Racewaytwitter.com · 2021-05-17

Where it's used

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.

No API provider serves it, so its downloads are people running it themselves: 41 quantized builds.

  • As a service0no API provider listed
  • On own hardware41quantized builds
  • Built on1313 fine-tunes · 0 adapters · 0 merges
  • In apps0Hugging Face Spaces

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 64 GB Mac (unified memory) at INT4

About 20 GB at INT4, 40 GB at FP8, 80 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 20 GB of weights (estimate; KV cache and activations extra)INT420 GBFP8: about 40 GB of weights (estimate; KV cache and activations extra)FP840 GBBF16: about 80 GB of weights (estimate; KV cache and activations extra)BF1680 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 20 GB of weights (estimate; KV cache and activations extra)INT420 GBFP8: about 40 GB of weights (estimate; KV cache and activations extra)FP840 GBBF16: about 80 GB of weights (estimate; KV cache and activations extra)BF1680 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2026-06-20), amber triangles are 10 dated articles (each links to it).

OctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctLaguna S 2.1 · 2026-07-21Laguna XS.2 and M.1 · 2026-04-28Laguna XS 2.1 33B on a RTX 3090: 296 tok/s peak, 152 tok/s at 256K context · 2026-07-11Anybody tried Laguna S 2.1 (by Poolside)? · 2026-08-06Laguna S 2.1:118B-a9B better than Qwen3.5:122B-a10B? So far, yes · 2026-08-06Agentic AI at Two Different Scales: Nanbeige4.2-3B and Laguna S2.1 · 2026-07-31Laguna XS.2 and Laguna M.1 by Poolside · 2026-04-28Laguna M.1/XS.2 Technical Report [pdf] · 2026-05-26Watch Tesla's Record-Breaking Laguna Seca Lap · 2019-09-12Laguna XS.2: A Coding Model Built for Closed Environments, Now Open Weights · 2026-04-29Weights 2026-06-20 · 2026-06-20Weights 2026-06-20now
OctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctLaguna S 2.1 · 2026-07-21Laguna XS.2 and M.1 · 2026-04-28Laguna XS 2.1 33B on a RTX 3090: 296 tok/s peak, 152 tok/s at 256K context · 2026-07-11Anybody tried Laguna S 2.1 (by Poolside)? · 2026-08-06Laguna S 2.1:118B-a9B better than Qwen3.5:122B-a10B? So far, yes · 2026-08-06Agentic AI at Two Different Scales: Nanbeige4.2-3B and Laguna S2.1 · 2026-07-31Laguna XS.2 and Laguna M.1 by Poolside · 2026-04-28Laguna M.1/XS.2 Technical Report [pdf] · 2026-05-26Watch Tesla's Record-Breaking Laguna Seca Lap · 2019-09-12Laguna XS.2: A Coding Model Built for Closed Environments, Now Open Weights · 2026-04-29Weights 2026-06-20 · 2026-06-20Weights 2026-06-20now

What people built on it

54 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

54 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

41 quantized (Hugging Face model tree)13 fine-tunes (Hugging Face model tree)41 quantized13 fine-tunes
41 quantized (Hugging Face model tree)13 fine-tunes (Hugging Face model tree)41 quantized13 fine-tunes
GGUF mudler/Laguna-XS-2.1-APEX-GGUF for llama.cpp, Ollama, LM Studio 5.2M downloads MLX mlx-community/Laguna-XS-2.1-4bit for Apple silicon 7.3k downloads NVFP4 poolside/Laguna-XS-2.1-NVFP4 for vLLM, SGLang 247.4k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

What the card says

1 cited sentence

Summary generated, cited

This model is released under the openmdw-1.1 license.[1]

  1. “license:openmdw-1.1”

Generated 2026-10-23 20:12 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

1 use case

What it is good at 1 use case

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.

Facts

16 fields

Listed in LLMs · MoE

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesLagunaForCausalLM
Model typelaguna
Params by dtypeBF16: 33,442,617,088
Size bandL
MoEyes
Experts256
Experts per token8
Hub snapshot2026-10-26

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Laguna-XS-2.1: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)80 GB · Multi-GPU node (<640 GB)
FP8 (1 byte/param)40 GB · Single high-end GPU (<80 GB)
INT4 (0.5 bytes/param)20 GB · Single GPU (<24 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

40 variants

Lineage

Original model Family: poolside/Laguna 54 derivatives on the Hub

quantized: 41finetune: 13

Top variants (40 shown, ranked by downloads)

Relation
Quant
Size
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
mudler/Laguna-XS-2.1-APEX-GGUF quantized GGUF — — openmdw-1.1 mudler 28 5.2M
poolside/Laguna-XS-2.1-GGUF quantized GGUF — — openmdw-1.1 poolside 100 646.4k
poolside/Laguna-XS-2.1-NVFP4 quantized NVFP4 33.4B L openmdw-1.1 poolside 16 247.4k
bartowski/Laguna-XS-2.1-GGUF quantized GGUF — — openmdw-1.1 bartowski 20 44.4k
poolside/Laguna-XS-2.1-FP8 quantized -FP8 33.4B L openmdw-1.1 poolside 9 24.6k
AtomicChat/Laguna-XS-2.1-GGUF quantized GGUF — — openmdw-1.1 AtomicChat 17 21k
poolside/Laguna-XS-2.1-INT4 quantized INT4 40.7B XL openmdw-1.1 poolside 6 19.4k
lmstudio-community/Laguna-XS-2.1-GGUF quantized GGUF — — openmdw-1.1 lmstudio-community 5 13k
ggml-org/Laguna-XS-2.1-GGUF quantized GGUF — — openmdw-1.1 ggml-org 4 9.7k
mradermacher/Laguna-XS-2.1-i1-GGUF quantized GGUF — — openmdw-1.1 mradermacher 2 8.9k
poolside/Laguna-XS-2.1-DFlash finetune — 462M S openmdw-1.1 poolside 15 8.2k
mlx-community/Laguna-XS-2.1-4bit quantized -4bit 33.4B L openmdw-1.1 mlx-community 4 7.3k
AtomicChat/Laguna-XS-2.1-MLX-8bit quantized MLX 33.4B L openmdw-1.1 AtomicChat 2 4.9k
poolside/Laguna-XS-2.1-NVFP4-mlx quantized NVFP4 33.4B L openmdw-1.1 poolside 3 4.7k
AtomicChat/Laguna-XS-2.1-MLX-5bit quantized MLX 33.4B L openmdw-1.1 AtomicChat 3 3.9k
dealignai/Laguna-XS-2.1-CRACK-GGUF quantized GGUF — — openmdw-1.1 dealignai 5 3.3k
Myric/Laguna-XS-2.1-APEX-GGUF quantized GGUF — — apache-2.0 Myric 0 3.3k
cyankiwi/Laguna-XS-2.1-AWQ-INT4 quantized AWQ 33.4B L openmdw-1.1 cyankiwi 3 3.1k
AtomicChat/Laguna-XS-2.1-MLX-6bit quantized MLX 33.4B L openmdw-1.1 AtomicChat 2 2.7k
mradermacher/Laguna-XS-2.1-GGUF quantized GGUF — — openmdw-1.1 mradermacher 0 2.8k
mlx-community/Laguna-XS-2.1-8bit quantized -8bit 33.4B L openmdw-1.1 mlx-community 2 2.1k
gbuzhf/Laguna-XS-2.1-APEX-GGUF quantized GGUF — — openmdw-1.1 gbuzhf 0 1.9k
mlx-community/Laguna-XS-2.1-OptiQ-4bit quantized -4bit 33.4B L openmdw-1.1 mlx-community 3 1.4k
noctrex/Laguna-XS-2.1-MXFP4_MOE-GGUF quantized GGUF — — — noctrex 0 1.4k
mlx-community/Laguna-XS-2.1-3bit quantized — 33.4B L openmdw-1.1 mlx-community 3 1.1k
mlx-community/Laguna-XS-2.1-bf16 finetune -bf16 33.4B L openmdw-1.1 mlx-community 2 1.2k
sahilchachra/Laguna-XS-2.1-W4A16 quantized -W4A16 33.7B L openmdw-1.1 sahilchachra 2 1.1k
leonsarmiento/Laguna-XS-2.1-6bit-XL-mlx quantized mlx 33.4B L — leonsarmiento 1 995
mlx-community/Laguna-XS-2.1-5bit quantized — 33.4B L openmdw-1.1 mlx-community 2 893
mlx-community/Laguna-XS-2.1-6bit quantized — 33.4B L openmdw-1.1 mlx-community 1 938
ji-farthing/Laguna-XS-2.1-DFlash-ik-llama-GGUF quantized GGUF — — openmdw-1.1 ji-farthing 4 776
dealignai/Laguna-XS-2.1-JANG_6M-CRACK finetune — 33.4B L other dealignai 0 924
tepirale/Laguna-S-2.1-DFLASH-GGUF quantized GGUF — — openmdw-1.1 tepirale 2 790
sahilchachra/Laguna-XS-2.1-NVFP4A16 quantized — 19.0B L openmdw-1.1 sahilchachra 1 753
dealignai/Laguna-XS-2.1-JANG_4M-CRACK finetune — 33.4B L other dealignai 0 780
FenomAI/Laguna-XS-2.1-AWQ-INT4 quantized AWQ 34.5B L openmdw-1.1 FenomAI 0 653
dealignai/Laguna-XS-2.1-JANG_2L-CRACK finetune — 33.4B L other dealignai 0 593
turboderp/Laguna-XS-2.1-exl3 quantized exl3 — — openmdw-1.1 turboderp 1 478
AxionML/Laguna-XS-2.1-NVFP4 quantized NVFP4 33.4B L other AxionML 0 502
OsaurusAI/Laguna-XS-2.1-JANG_6M finetune — 33.4B L other OsaurusAI 0 455

Repository changelog

1 change
changed in the last 7 daysNew on the watch list: 2026-10-23all changes

Articles and coverage

12 pieces

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.

Analysis8

Laguna S 2.1
poolside.ai2026-07-21HN front pageabout the Laguna familyHN 416 pts · 88 comments3,843 words
Poolside releases Laguna S 2.1, an agentic coding model with 118B total parameters and 8B activated parameters per token that launched in under nine weeks. It supports a context window of up to 1M tokens, scores 70.2% on Terminal-Bench 2.1, and has a knowledge cutoff date in November 2025. Full evaluation trajectories are available for every trial in the final evaluation set.written, reviewed, cited
  • Laguna S 2.1 scored 40.4% in our agent harness, pool, not mini-swe-agent which DeepSWE's leaderboard uses.
Laguna XS.2 and M.1
poolside.ai2026-04-28HN front pageabout the Laguna familyHN 101 pts · 50 comments3,078 words
Poolside AI releases Laguna XS.2 and Laguna M.1, agentic coding models now served at 256K context lengths with stronger scores on SWE-bench Verified and Terminal-Bench 2.0. Laguna XS.2 is a 33B total parameter MoE with 3B activated parameters, trained on 30T tokens using NVIDIA Hopper GPUs, reaching 44.5% on SWE-Bench Pro and 30.1% on Terminal-Bench 2.0. The model is available under an Apache 2.0 license and free via API and OpenRouter for a limited time.written, reviewed, cited
  • Benchmarking used a maximum of 500 steps and sandboxed execution using 8 GB RAM/2 CPUs for most tasks.
  • The Laguna series uses approx. 4.4T+ synthetic tokens to complement natural web data throughout pre-training.
  • Laguna XS.2 is supported in NVIDIA TensorRT-LLM on Day 1 with an NVFP4 version for Blackwell architecture.
Laguna XS 2.1 33B on a RTX 3090: 296 tok/s peak, 152 tok/s at 256K context
lucebox.com2026-07-11HN 20 pts · 1 comments1,373 words
Laguna XS 2.1 33B runs on a single RTX 3090 with speculative decoding to achieve 296 tok/s peak speed and maintain 152 tok/s at 256K context. The system uses KVFlash paging with an 8K pool and ring caches to handle contexts that do not physically fit in 24 GB of VRAM while keeping output quality lossless. Prefill processes 256K tokens in 67 seconds at ~3,500 tok/s using a Qwen3-0.6B drafter for scoring KV chunks.written, reviewed, cited
  • The full KV cache for 256K tokens requires approximately 20 GB, exceeding the 24 GB VRAM limit of the RTX 3090.
  • Sliding-window layers skip the pager entirely by attending a 2K position ring instead of the whole KVFlash pool.
  • The prefill drafter compresses prompts and scores KV chunk residency, falling back to recency-only LRU without it.
Agentic AI at Two Different Scales: Nanbeige4.2-3B and Laguna S2.1
kaitchup.substack.com2026-07-31about the Laguna familyHN 2 pts · 0 comments1,362 words
Nanbeige4.2-3B and Laguna S 2.1 target agentic workloads at different scales. Nanbeige4.2-3B uses a Looped Transformer with 22 physical decoder layers executed twice to achieve 44 layer executions without storing 44 independent layers, fitting on a 16GB GPU but consuming roughly 44 GiB of KV-cache memory at 256K context. Laguna S 2.1 is a 118-billion-parameter Mixture-of-Experts model activating approximately 8 billion parameters per token, featuring 48 transformer layers and 256 routed experts, requiring vLLM 0.25.0 and a 96 GB GPU to exploit its 256K context length.written, reviewed, cited
  • Nanbeige's default configuration does not enable loop-level KV sharing, potentially requiring separate key/value entries for each pass.
  • Retaining reasoning traces in conversation history risks reaching the model's maximum context length on the very next turn.
Laguna XS.2 and Laguna M.1 by Poolside
poolside.ai2026-04-28about the Laguna familyHN 5 pts · 0 comments999 words
Laguna M.1/XS.2 Technical Report [pdf]
poolside.ai2026-05-26about the Laguna familyHN 2 pts · 0 comments18,357 words
Watch Tesla's Record-Breaking Laguna Seca Lap
engadget.com2019-09-12about the Laguna familyHN 6 pts · 2 comments380 words
Laguna XS.2: A Coding Model Built for Closed Environments, Now Open Weights
firethering.com2026-04-29about the Laguna familyHN 1 pts · 0 comments1,335 words

Community2

Anybody tried Laguna S 2.1 (by Poolside)?
news.ycombinator.com2026-08-06about the Laguna familyHN 7 pts · 2 comments113 words
Laguna S 2.1:118B-a9B better than Qwen3.5:122B-a10B? So far, yes
news.ycombinator.com2026-08-06about the Laguna familyHN 4 pts · 2 comments157 words
This American (San Francisco) model, Laguna S 2.1:118B-a9B, outperforms Qwen3.5:122B-a10B and Sonnet in current benchmarking tests. It demonstrates better personality and natural creativity when designing UIs. Benchmarks need to include a minimum of hardware requirements, which are currently missing.written, reviewed, cited
All coverage tracked for this model (12) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Laguna S 2.1:118B-a9B better than Qwen3.5:122B-a10B? So far, yesnews.ycombinator.com2026-08-062026-10-23hn
Anybody tried Laguna S 2.1 (by Poolside)?news.ycombinator.com2026-08-062026-10-23hn
Agentic AI at Two Different Scales: Nanbeige4.2-3B and Laguna S2.1kaitchup.substack.com2026-07-312026-10-23hn
Laguna S 2.1poolside.ai2026-07-212026-10-23hn
Laguna XS 2.1 33B on a RTX 3090: 296 tok/s peak, 152 tok/s at 256K contextlucebox.com2026-07-112026-10-23hn
Laguna M.1/XS.2 Technical Report [pdf]poolside.ai2026-05-262026-10-23hn
Laguna XS.2: A Coding Model Built for Closed Environments, Now Open Weightsfirethering.com2026-04-292026-10-23hn
Laguna XS.2 and Laguna M.1 by Poolsidepoolside.ai2026-04-282026-10-23hn
Laguna XS.2 and M.1poolside.ai2026-04-282026-10-23hn
Tesla is attempting lap record with a Model S Plaid+ at Laguna Seca Racewaytwitter.com2021-05-172026-10-23hn
Watch Tesla's Record-Breaking Laguna Seca Lapengadget.com2019-09-122026-10-23hn
Tesla Model S at Refuel 2013 (Laguna Seca)youtube.com2013-09-192026-10-23hn

Similar models

3 models

Similar models

  • GLM-4.7-Flash · 31.2B
    same size band (L), also MoE, same task (text-generation), more providers, permissive license
  • Qwen3-Coder-30B-A3B-Instruct · 30.5B
    same size band (L), also MoE, same task (text-generation), more providers, permissive license
  • Qwen3-30B-A3B · 30.5B
    same size band (L), also MoE, same task (text-generation), more providers, permissive license

Sources

4 sources

Sources

Hub snapshot: 2026-10-26 · Providers: 2026-10-26 07:26 UTC · Articles: 2026-10-26 07:27 UTC · Summary: 2026-10-23 20:12 UTC · Last verified 2026-10-26 07:27 UTC

Follow Laguna-XS-2.1

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.