Qwen3-Coder-Next-Spark-Agentic

Parameters

Total	~80B (HF count 79.67B BF16; non-embedding ~79B)
Active / token	~3B
Arch	MoE · Qwen3NextForCausalLM · 512 experts · 10 act · 1 shared
Source	Qwen/Qwen3-Coder-Next
This pack	runtime / probe only — no weights here

HF sidebar Parameters badge stays empty on purpose (no checkpoint in this repo). Use the table above for size. Full detail: Model parameters.

Preview / Unverified DGX Spark deployment pack for Qwen/Qwen3-Coder-Next (~80B total / ~3B active).

This repository is a runtime pack, not a model checkpoint and not a re-quant. Harness probe passed (2/2). Not a ship / gate claim. Verifier ≠ gate clearance.

Status

Public role	Preview · Probe-measured — no ship claim
Aug 3 role	Co-ship with Laguna S (same calendar)
Flagship measured	Laguna S remains sole flagship measured pack
Primary OS	Official FP8 + vLLM · tool parser qwen3_coder
GGUF alts	Official Q4_K_M (pulled) · optional Q5_K_M · llama.cpp
diy_gguf	false
agent_smoke	Probe 2/2 pass (2026-07-30) — not gate
public_promo_before_launch	false until 2026-08-03 12:00 WIB
Ports	Laguna :8000 · this pack default :8001
Tool format note	Qwen3-Coder native tool-call format — not Hermes JSON
Affiliation	Independent · personal · not Alibaba · not Ainfera product

Sibling flagship (measured): Laguna-S-2.1-Spark-Agentic

Launch calendar: LAUNCH_AUG3.md · lock: results/launch_lock.json

---

Measure track

Same closed harness as Laguna S after tools actually execute.

1. Pull FP8 → serve vLLM on :8001 ✅
2. Point eval/agent_smoke at http://127.0.0.1:8001/v1 ✅
3. Write receipt under results/ ✅ — no ship claim without dated gate
4. Verifier ≠ gate clearance ✅

Latest measure window: results/measure_window_fp8_20260730T114320+0700.json Smoke: results/agent_smoke.json · Receipt index: results/RECEIPT_INDEX.json

GGUF alts (llama.cpp) remain secondary headroom/quality paths.

---

Upstream

Field	Value
Base	Qwen/Qwen3-Coder-Next
Official FP8 (primary)	Qwen/Qwen3-Coder-Next-FP8
Official GGUF (alts)	Qwen/Qwen3-Coder-Next-GGUF
License	Apache-2.0 (full text in LICENSE)
Arch	Qwen3NextForCausalLM / qwen3_next

---

Model tree

Qwen/Qwen3-Coder-Next                 ← base (official)
├── Qwen/Qwen3-Coder-Next-FP8         ← primary OS weights
├── Qwen/Qwen3-Coder-Next-GGUF        ← official GGUF alts
└── hizrianraz/Qwen3-Coder-Next-Spark-Agentic  ← this pack (runtime · probe)

HF deploy-pack cards omit YAML base_model so Hub does not attach base_model:finetune / quantized tree relation. Upstream cited in prose + links only. This pack is not a checkpoint. No library_name: gguf. No base_model_relation: quantized. Probe ≠ gate.

---

Model parameters (upstream · not this pack)

This repo hosts no checkpoint weights, so the HF sidebar Parameters widget stays empty by design. Sizes below are copied from the upstream base — not re-counted here. Probe ≠ gate.

Field	Value	Source
Architecture	MoE · Qwen3NextForCausalLM / qwen3_next	Qwen config
Total parameters	~80B (HF safetensors count 79.67B BF16; non-embedding ~79B)	Qwen/Qwen3-Coder-Next
Activated / token	~3B	upstream card
Experts	512 total · 10 activated · 1 shared · expert intermediate 512	upstream card + config
Pack role	runtime / probe harness only	this repo

---

Download weights (official only)

# on Spark — primary OS
chmod +x scripts/*.sh
./scripts/pull_official_fp8.sh
# expected:
#   ~/models/qwen3-coder-next/Qwen3-Coder-Next-FP8/*.safetensors

# GGUF alts (secondary)
./scripts/pull_official_gguf.sh
# expected:
#   ~/models/qwen3-coder-next/Qwen3-Coder-Next-Q4_K_M/*.gguf

Q4_K_M shard pins (Spark pull 2026-07-29)

Shard	SHA256	Bytes
...-00001-of-00004.gguf	6bcfc9f9c37901eeb92172e2ab871224dab36a453d263bcb2547f737409534da	15524827040
...-00002-of-00004.gguf	817def0691ee9d08bf3dc4444be7aed29c9e52091e8fa9d97901ce7e7f6f01d3	14872168352
...-00003-of-00004.gguf	23aa634d47dca9b4ca3ea249384e6f01951b24c83cdc076f37f6f43d6c99883f	14503294496
...-00004-of-00004.gguf	249c768cc5f130dc731567d6edcbdacc48e14dec9e02c5dbe2b2185d2c5bdb2b	3510702144

See docs/WEIGHTS_LOCAL.md + SHA256SUMS. Prefer revision pins over floating main.

---

---

Safe quickstart (probe only)

1. Leave Laguna on :8000. This pack defaults to :8001.
2. Prefer FP8 + vLLM OS path: ./scripts/serve_vllm_fp8.sh (binds host 127.0.0.1 unless you override HOST).
3. Smoke always writes a receipt under results/ — probe ≠ gate clearance.
4. Do not set PORT=8000 without MEASURE_WINDOW=1.
5. No ship / hero CTA from this card until an explicit measured gate — tag stays unmeasured / probe-measured.

# host loopback defaults; override only when you mean it
./scripts/serve_vllm_fp8.sh
OPENAI_BASE_URL=http://127.0.0.1:8001/v1 \
SMOKE_MODEL=local-qwen3-coder-next \
python3 eval/agent_smoke/run_smoke.py
# receipt: results/agent_smoke.json

Serve (measure window)

chmod +x scripts/*.sh
# PRIMARY OS — FP8 + vLLM (default :8001; leave Laguna on :8000)
./scripts/serve_vllm_fp8.sh

# GGUF alt — llama.cpp (also :8001; stop vLLM first)
./scripts/serve_spark.sh

Do not set PORT=8000 without MEASURE_WINDOW=1.

---

Measurement bar

Same integrity rules as Laguna packs:

• smoke receipts under results/ before any pass claim
• verifier ≠ gate clearance
• no “agentic model” marketing from unrun packs

OPENAI_BASE_URL=http://127.0.0.1:8001/v1 \
SMOKE_MODEL=local-qwen3-coder-next \
python3 eval/agent_smoke/run_smoke.py

Current measure state: probe_measured_not_gate — FP8 complete + measure-window smoke 2/2; ship claim still forbidden until explicit gate.

---

Status ledger

Gate	State
Scaffold	done
Official Q4_K_M pull (Spark)	complete (digests pinned)
Official FP8 pull (Spark)	complete 40/40 + SHA (results/fp8_sha256_latest.json)
vLLM serve script	wired (scripts/serve_vllm_fp8.sh)
measure window	pass 2026-07-30 11:53 WIB · Laguna restored
agent_smoke	probe 2/2 pass · not gate
Public promo before 08-03 12:00 WIB	forbidden
Classification	deployment pack · not a model checkpoint

---

Client surface

OpenAI-compatible notes + sample client: hermes/ (Qwen native tool format — not Hermes JSON.)

Related (Aug 3 family)

• Flagship measured: Laguna-S-2.1-Spark-Agentic · 40/40
• Co-list scaffold (unmeasured): DeepSeek-V4-Flash-REAP25-Spark-Agentic · no ship claim (legacy DeepSeek-V4-Flash-Spark-Agentic → redirect)
• Laguna XS Mac — parked until Metal MoE path is trustworthy + measured
• Qwen3-Coder-30B Mac — parked until measured

Family matrix

See parent FAMILY_MATRIX_AUG3.md + results/triple_aug3_lock.json.

Updated: 2026-07-30 14:10 WIB — safe quickstart + omit base_model YAML + unmeasured tags