Laguna-S-2.1-Spark-Agentic

Measured DGX Spark deployment pack for poolside/Laguna-S-2.1 (118B total · ~8B active/token · MoE · agentic coding).

This repository is a runtime / evaluation pack, not a model checkpoint and not a re-quant. Hugging Face may still list it under Models; treat it as a runbook + harness + receipts.

Positioning

Unique contribution: proving how the official Poolside Q4_K_M behaves on a single DGX Spark — pinned engine, fail-closed digests, closed harness, honest negatives (DFlash).

Not coded as: boutique quant producer, base-model trainer, or long-horizon agent reliability proof.

Role	Flagship measured deployment recipe
Weights authority	poolside/Laguna-S-2.1-GGUF
Optional mirror	hizrianraz/Laguna-S-2.1-GGUF · byte-identical upstream · not re-quantized by this account
Default quant	Official Poolside Q4_K_M · ~89.4 GiB
Host	NVIDIA DGX Spark (GB10)
agent_smoke	40/40 tool-format / routing regression smoke (see honesty section)
Launch window	2026-08-03 12:00 WIB · freeze 2026-08-02 18:00 WIB
Affiliation	Independent · not Poolside · not Nous Research

Launch calendar: LAUNCH_AUG3.md

---

Download weights (authoritative = Poolside)

Prefer upstream; mirror is convenience only:

# recommended
hf download poolside/Laguna-S-2.1-GGUF laguna-s-2.1-Q4_K_M.gguf --local-dir ~/models/laguna-s-2.1

# optional convenience mirror (same bytes)
hf download hizrianraz/Laguna-S-2.1-GGUF laguna-s-2.1-Q4_K_M.gguf --local-dir ~/models/laguna-s-2.1

echo "a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4  laguna-s-2.1-Q4_K_M.gguf" \
  | (cd ~/models/laguna-s-2.1 && sha256sum -c -)

This pack is not the bulk weight host. It publishes serve pins, digests, and smoke harness.

---

Why this pack

1. Pinned Spark serve settings that clear tool-format smoke
2. Fixed harnesses you can re-run (no vibes)
3. Official digests for fail-closed verify
4. OpenAI-compatible client notes for tool-agent stacks
5. Negative results retained (DFlash slower — do not promote)

---

Highlights (measured tip · 2026-07-29 13:22 WIB)

Metric	Result
Host	NVIDIA DGX Spark (GB10)
Quant	Official laguna-s-2.1-Q4_K_M.gguf
Engine	poolsideai/llama.cpp @ 04b2b72 (branch laguna)
Serve context flag	-c 8192
agent_smoke	40/40 · 84.86 s · temp 0.0 · format/routing smoke
hermes_agent_smoke v2	27/27 · 100.1 s · temp 0.0 · one-response; tools validated not executed
Generation throughput (sole short-gen headline)	~21.47 tok/s @ 128 completion tokens (prompt 67)
Measure tip	pack bf82eab

Artifacts: results/MEASURED.md · results/measured.json · results/agent_smoke.json · results/hermes_agent_smoke.json

---

Honesty clock (read before quoting scores)

What 40/40 and 27/27 are

Useful tool-format and routing regression smokes on a fixed case list.

Receipt language (hermes suite): tools validated, not executed.

What they are not

• Not long-horizon agent reliability
• Not multi-turn tool-result continuation proof
• Not exact argument-value / schema-hard proof for every case
• Not SWE-bench / BFCL / Terminal-Bench substitutes

Bench label correction (server_bench marks)

Recorded prefill marks were labeled 2k / 8k for convenience. Actual prompt token counts on the receipt:

Mark label	Actual prompt_tokens	Latency	Completion
2k	836	1.597 s	3 tokens ("OK")
8k	3236	4.78 s	3 tokens ("OK")

Do not cite these as full 2K/8K filled-context prefill rates.

Generation headline limits

• ~21.47 tok/s is one short-gen point: 67 prompt / 128 completion tokens
• Same-day suite also logged gen8 ~10.6 t/s and gen256 ~21.9 t/s
• Not a multi-seed soaked long-session figure; warm-rep / p50-p95 TTFT / peak RAM still to strengthen

External signal bar (future, not claimed): BFCL V4 · τ-bench · Terminal-Bench 2.

---

Quick start

# 1) Engine — pin the measured commit
git clone https://github.com/poolsideai/llama.cpp
cd llama.cpp
git checkout 04b2b72cb54048ead292884adbe11f284e3ec950
# Spark/GNU may need the isfinite patch in docs/BUILD_SPARK.md
cmake -B build -G "Unix Makefiles" \
  -DCMAKE_BUILD_TYPE=Release \
  -DGGML_CUDA=ON \
  -DCMAKE_CUDA_ARCHITECTURES=121a \
  -DLLAMA_CURL=ON
cmake --build build -j --target llama-server llama-cli llama-bench

# 2) Weights — upstream first + fail-closed digest
huggingface-cli download poolside/Laguna-S-2.1-GGUF \
  laguna-s-2.1-Q4_K_M.gguf \
  --local-dir ~/models/laguna-s-2.1
cp /path/to/this/pack/SHA256SUMS ~/models/laguna-s-2.1/
(cd ~/models/laguna-s-2.1 && sha256sum -c SHA256SUMS)

# 3) Serve (flags match last-green pin)
./build/bin/llama-server \
  -m ~/models/laguna-s-2.1/laguna-s-2.1-Q4_K_M.gguf \
  --host 127.0.0.1 --port 8000 \
  --ctx-size 8192 -ngl -1 --jinja \
  -fa on --alias local-laguna

# 4) Smoke — model id must match --alias
cd /path/to/this/pack
python eval/agent_smoke/run_smoke.py \
  --base-url http://127.0.0.1:8000/v1 \
  --model local-laguna

# 5) Optional hermes-class suite (does not replace step 4)
python eval/hermes_agent_smoke/run_hermes_smoke.py \
  --base-url http://127.0.0.1:8000/v1 \
  --model local-laguna \
  --out results/hermes_agent_smoke.json

Guide	Path
Spark notes	SPARK.md
Reproduce	docs/REPRODUCE.md
Build	docs/BUILD_SPARK.md
Last-green pin	results/LAST_GREEN_PIN.md
Sample client	hermes/

---

License

Base model and official GGUF: OpenMDW-1.1 (Poolside).

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

Retain notices. Pack scripts and eval harnesses are separate files with clear provenance.

---

Base identity (pinned)

Field	Value
Base model	poolside/Laguna-S-2.1
Base revision	00af5a51782109b587a3b3bbf11875e566036fa7
Architecture	118B total · ~8B active/token · MoE
Official GGUF (authority)	poolside/Laguna-S-2.1-GGUF
Optional mirror	hizrianraz/Laguna-S-2.1-GGUF · mirror only
GGUF revision (upstream pin)	fc4e481289523cf7d0df668da6d1d391616141ca
Stand-behind quant	official laguna-s-2.1-Q4_K_M.gguf
Engine	poolsideai/llama.cpp · branch laguna · PR #25165

Official GGUF digests (HF LFS metadata)

File	sha256
laguna-s-2.1-Q4_K_M.gguf	a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4
laguna-s-2.1-Q8_0.gguf	d946b221d69f2c5f87a986952bcd3cfb75831e5a6a2184e626e361663e1bfe2b
laguna-s-2.1-F16.gguf	2036c9dcf70f59738d480d1da2a2a59c0a3a3c5bd4ab8c43a61d77fba031e1f8
laguna-s-2.1-DFlash-BF16.gguf	2ee8aa30338d6599bc7a8ce008cc57c56f2c2b2fdc21f6db9ecda203c751bfd4
laguna-s-2.1.imatrix	4a4f480f57a3251e3acfb1d35ffba64720662536135e4ca4f4d05b0732539be2

Pack verify file (GNU 2-column, Q4 default): SHA256SUMS

---

Which quant to pull

Goal	Artifact	Notes
Default agent serve (Spark)	Poolside S Q4_K_M	Headline smoke + short-gen
Higher fidelity (Spark)	Poolside S Q8_0	More RAM
Speculative decode experiment	+ Poolside DFlash-BF16	Measured slower · DO_NOT_PROMOTE
Full S on Mac ≤32 GB RAM	No local S weights	Use Mac as client → Spark only

diy_gguf: false. Do not claim “first quant”. Mirror ≠ re-quant.

---

Tool-agent wiring (OpenAI-compatible)

OpenAI chat.completions + tools / tool_calls shape. Not a Nous Research product claim or endorsement. “Hermes-class” in older notes = that wire shape only.

export OPENAI_BASE_URL=http://127.0.0.1:8000/v1   # /v1 required
export OPENAI_API_KEY=sk-local
export OPENAI_MODEL=local-laguna

python hermes/sample_client.py

python eval/agent_smoke/run_smoke.py \
  --base-url "$OPENAI_BASE_URL" --model local-laguna
python eval/hermes_agent_smoke/run_hermes_smoke.py \
  --base-url "$OPENAI_BASE_URL" --model local-laguna \
  --out results/hermes_agent_smoke.json

Agent loop contract (production agents; richer than current smoke):

1. POST messages + tool schemas
2. If tool_calls → execute only offered names → append role=tool
3. Repeat until final content
4. Sanitize invalid prior function.arguments JSON before re-send

Config snippet: hermes/config.example.yaml

---

Scoreboard (honest)

Freeze clock: 2026-08-02 18:00 WIB — no new claims after freeze without re-measure.

Host	Quant	Ctx flag	Gen tok/s (128 compl.)	agent_smoke	hermes v2	Role
DGX Spark GB10	official Q4_K_M	8192	~21.47 (short-gen)	40/40 format smoke	27/27 validated-not-executed	Headline pack · engine 04b2b72 · tip bf82eab

Ship minimum smoke ceiling historically used: ≥38/40 on this harness. Current: 40/40.

Mandate: strengthen with multi-seed, actual tool execution, arg-value checks, warm TTFT stats — not ship day scope unless left green after freeze.

---

Scope

In scope

• Measured Spark deployment packing
• Official weight pointers + digests
• Reproducible smoke harness + receipts

Out of scope

• Claiming this account quantized the weights
• Classifying this docs pack as a loadable GGUF model in prose
• Long-horizon agent SOTA claims from format smoke alone
• Fake or off-host benches labeled as Spark
• Local full-S weights on Mac hosts with ≤32 GB RAM
• Public announce / trending push before 2026-08-03 12:00 WIB
• Shipping unmeasured XS/Qwen30/DeepSeek as peer “models”

---

Attribution

Component	Credit
Model	Poolside Laguna S 2.1 © Poolside · OpenMDW-1.1
GGUF	Poolside official conversions (authoritative)
Optional mirror	Byte-identical copy for bandwidth convenience only
Engine	poolsideai/llama.cpp laguna (+ upstream llama.cpp)
Pack, measurements, smokes	Independent work by hizrianraz

Disclaimer

Independent measurements on one DGX Spark. Not affiliated with, endorsed by, or representing Poolside or Nous Research.

Updated local stage (Sol reframe): 2026-07-30 00:45 WIB