Laguna-XS-2.1-Mac-Agentic

Mac-class research pack for poolside/Laguna-XS-2.1 (33B total · ~3B active/token · MoE).

Status	Research / prep surface
Measured smoke	None yet
Ship claim	None
Default disk candidate	Official Q4_K_M (~18.9 GiB)
Weights (default)	hizrianraz/Laguna-XS-2.1-GGUF
Not	A quant of Laguna-S · “S on Mac” · S-lite
Affiliation	Independent · not Poolside · not Nous Research

Download weights here: hizrianraz/Laguna-XS-2.1-GGUF (Laguna-XS-2.1-Q4_K_M.gguf · official Poolside bytes · sha256-matched)

This is a separate model family from Laguna-S. Full-S weights and headline measurements live only on the sibling pack:

Laguna-S-2.1-Spark-Agentic (S freeze 2026-08-02 18:00 WIB · S launch 2026-08-03 12:00 WIB)

XS does not couple to the S freeze or S launch clock.

---

Bottom line

Field	Value
Model	Laguna-XS-2.1 · 33B-A3B MoE
Host class	Apple Silicon Mac with ≤32 GB unified memory
Disk candidate	Official Laguna-XS-2.1-Q4_K_M.gguf · ~18.9 GiB
Runtime	Unproven on ≤32 GB Mac hosts
agent_smoke / tok/s	Not measured — do not invent
Full S on this host class	Non-fit local — use S pack as client → Spark

Disk-fit ≠ load-fit ≠ agent-smoke-fit.

---

What this pack provides

1. Honest Mac envelope for XS authentic (not “S on Mac”)
2. Official GGUF pointers and sha256 — default download via hizrianraz/Laguna-XS-2.1-GGUF
3. Shared agent / Hermes smoke harness shape (empty results until a real run)
4. Pull + Metal serve scaffolds

What this pack is not

• Not Laguna-S-2.1
• Not “S-lite”, “S-Mac”, or “S via XS”
• Not a measured pass-rate or throughput claim
• Not a bulk GGUF mirror
• Not the Aug 3 S launch headline

---

License

Base model and derivatives: OpenMDW-1.1 (Poolside).

• Pack: LICENSE.md
• Upstream: https://huggingface.co/poolside/Laguna-XS-2.1/blob/main/LICENSE.md

Retain notices. Pack scripts/eval provenance: NOTICE.

---

Base identity (pinned)

Field	Value
Base model	poolside/Laguna-XS-2.1
Base revision	205dc65dd4bda946c50da6b7522b215734fa107b
Architecture	33B total · ~3B active/token · MoE
Official GGUF repo (upstream)	poolside/Laguna-XS-2.1-GGUF
Preferred mirror	hizrianraz/Laguna-XS-2.1-GGUF
GGUF revision (upstream pin)	1a37c0a5fb8c7a18e6106decb6be6327d1b63fa6
Disk candidate	Q4_K_M · 20,274,300,032 bytes · 18.882 GiB
Non-fit on ≤32 GB	BF16 GGUF ~62 GiB · safetensors base ~67G class
Engine	Prefer poolsideai/llama.cpp branch laguna (Metal) · PR #25165
Stock Homebrew llama	Do not assume

Official GGUF digests

File	sha256	≤32 GB Mac
Laguna-XS-2.1-Q4_K_M.gguf	1ac7079101fca5a6df8c5a7523a3c30ea7d1c0e4b1258090e7d6d4039287f6cb	Disk candidate only
Laguna-XS-2.1-BF16.gguf	d1d99abe30c37749ec1b1ae2681cec74caa550ac6243d2ea70b61b1ff9d187ca	Non-fit

---

Mac host envelope (≤32 GB class)

Fact	Value
Target class	Apple Silicon Mac · ≤32 GB unified memory
Vendor language caution	Cards that say “Mac with 36 GB” are a looser envelope than 32 GB
Preferred weight volume	External SSD — avoid filling System Data
Full S Q4 (~96G class)	Non-fit local

For full S, use a Mac as a client only:

export OPENAI_BASE_URL=http://<spark-host>:8000/v1
export OPENAI_API_KEY=sk-local
export OPENAI_MODEL=local-laguna

See the S pack for measured Spark numbers.

---

Quick start

Weights are not pulled into this pack by default. Do not pull BF16 on ≤32 GB hosts.

# 1) Pull ~19G Q4 only — prefer an external volume
./scripts/pull_xs_q4.sh "/path/to/external/models/laguna-xs-2.1"

# 2) Verify
shasum -a 256 "/path/to/external/models/laguna-xs-2.1/Laguna-XS-2.1-Q4_K_M.gguf"
# expect:
# 1ac7079101fca5a6df8c5a7523a3c30ea7d1c0e4b1258090e7d6d4039287f6cb

# 3) Serve (requires Laguna-capable llama-server Metal build)
./scripts/serve_mac.sh \
  "/path/to/external/models/laguna-xs-2.1/Laguna-XS-2.1-Q4_K_M.gguf"

# 4) Smoke — only after load proof
python3 eval/agent_smoke/run_smoke.py \
  --base-url http://127.0.0.1:8000/v1

---

Measurement gate (before any public claim)

1. Working Laguna-capable engine on the target Mac — record commit
2. sha256 matches lock
3. Load proof: RSS + free RAM after load · context length
4. Same eval/agent_smoke 40 (or named subset + explicit disclaimer)
5. Optional hermes-class 27
6. Host labeled as Mac ≤32 GB (never as Spark)
7. Fail list + separate scoreboard row (never dilute S headline)
8. Explicit go before any public claim move

Results belong under results/ when real. Empty by design today.

---

Repository layout

Path	Role
MAC.md	Mac Metal serve envelope
docs/REPRODUCE.md	Reproduce steps
docs/BUILD_MAC.md	llama.cpp Laguna Metal build notes
eval/	agent + hermes smoke harness
hermes/	OpenAI-compatible sample client
scripts/pull_xs_q4.sh	Official Q4 pull + sha check
scripts/serve_mac.sh	Local Metal serve scaffold
results/launch_lock.json	Prep lock (B-track)
research/	Fit notes + dual roadmap

---

Dual roadmap

Track	Surface	Weight host
A · S	Laguna-S-2.1-Spark-Agentic	DGX Spark only
B · XS	This repository	Mac ≤32 GB candidate

Notes: research/roadmap-dual-s-and-xs.md

---

Honesty rules

Allowed	Forbidden
“XS Q4 is the official ~19G Mac-class GGUF for Laguna-XS-2.1”	“Laguna runs on Mac” (unqualified)
“Mac client still uses Spark for full S”	Any unmeasured pass rate or t/s
“Disk candidate only until load + smoke”	“S on Mac via XS” / S-lite / S-Mac
“Separate 33B-A3B model”	BF16 / safetensors as Mac-local ship path

---

Status

Field	Value
Lock	results/launch_lock.json → prep_published
Weights in repo	false
Measured	false
May delay S freeze	false

Created 2026-07-28. Prepared research parallel only. No weight pull in-tree. No smoke. No ship claim.

Attribution

Component	Credit
Model	Poolside Laguna XS 2.1 © Poolside · OpenMDW-1.1
GGUF	Poolside official conversions
Engine	poolsideai/llama.cpp laguna (+ upstream llama.cpp)
Pack / harness scaffolding	Independent work by hizrianraz

Disclaimer

Independent research pack. Not affiliated with, endorsed by, or representing Poolside or Nous Research. “Hermes-class” describes an OpenAI-compatible tool-calling runtime shape only — not a Nous endorsement.