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
Upstream native context	262,144 tokens
This launch profile	32,768-token configured ceiling; effective context unmeasured
Day-0 artifact	Official Qwen3-Coder-Next-FP8
Source	Qwen/Qwen3-Coder-Next
This pack	deployment / evaluation pack only — no weights here

HF native Parameters badge stays empty (no checkpoint in this repo) — size is the table above. Native Model tree is enabled via YAML base_model: Qwen/Qwen3-Coder-Next for findability only — this is not a finetune/adapter/quant of that checkpoint.

What is ready

Component	Audited day-0 choice
Checkpoint	Official FP8, revision da6e2ed27304dd39abadd9c82ef50e8de67bdd4c
Integrity	Complete 52-file snapshot bound by SHA-256, including 40/40 shards
Runtime	vLLM 0.25.1, immutable OCI index and explicit linux/arm64 selection
Spark profile	TP=1 · 32K context ceiling · one sequence · memory utilization ≤0.80
Endpoint	OpenAI-compatible · host-published 127.0.0.1:8001/v1 · /v1 key required
Tool parser	Native qwen3_coder parser
Residency	Single-model policy; automated known-process/container, memory, and swap screen
Public class	Preview; future SAQS certification is a separate process

The canonical machine-readable deployment manifest is config.yaml. INSTALL.yaml and hermes/config.yaml are discovery/compatibility pointers only.

Requirements

• DGX Spark with 128 GB unified memory
• A Git checkout pinned to the published 40-hex release commit
• Local Docker default context on /var/run/docker.sock plus NVIDIA Container Toolkit
• A reviewed Hugging Face CLI executable bound by its absolute path and SHA-256
• A 32–256 character random VLLM_API_KEY
• At least 96 GiB of available memory before loading; no active swap
• Roughly 90 GB of free storage for the official FP8 snapshot and download overhead
• No Laguna, DeepSeek, llama.cpp, vLLM, SGLang, or other model server resident

The launcher fails closed on known model-server processes/containers, insufficient available memory, and active swap. That screen does not enumerate every GPU compute process, so the freeze operator must separately attest that the host is quiescent. Using a different port does not make two large models safely co-resident.

Install the pack without a floating source

Use the exact 40-hex commit recorded by the authorized release. main, a branch, a moving tag, and a self-derived “current HEAD” are not source-authority pins:

git clone https://github.com/hizrianraz/Qwen3-Coder-Next-Spark-Agentic.git
cd Qwen3-Coder-Next-Spark-Agentic
PUBLISHED_COMMIT='REPLACE_WITH_AUTHORIZED_40_HEX_COMMIT'
git checkout --detach "$PUBLISHED_COMMIT"

export QWEN_ALLOWED_PACK_ROOT="$PWD"
export QWEN_EXPECT_PACK_REVISION="$PUBLISHED_COMMIT"
install -d -m 0700 "$HOME/models/qwen3-coder-next"
export QWEN_ALLOWED_MODEL_ROOT="$HOME/models/qwen3-coder-next"

Review config.yaml, then download and verify the official FP8 weights. Bind the reviewed hf executable before use. The puller uses an exact upstream revision, a new private staging directory, 52 hashes, and no-overwrite publication; it removes proxy and Hub endpoint overrides and gives the downloader a fresh private home instead of exposing the operator's home configuration.

export QWEN_HF_BIN=/absolute/reviewed/path/to/hf
export QWEN_HF_BIN_SHA256='REPLACE_WITH_REVIEWED_64_HEX_SHA256'
./scripts/pull_official_fp8.sh

# Pre-pull the exact arm64 image; the launcher itself uses --pull never.
docker pull --platform linux/arm64 \
  vllm/vllm-openai@sha256:e4f88a835143cd22aee2397a26ec6bb80b3a4a6fe0c882bcbc63822904766089
export VLLM_API_KEY='REPLACE_ME'  # Deliberately too short; replace before launch.
./scripts/serve_vllm_fp8.sh

The only accepted model path is $QWEN_ALLOWED_MODEL_ROOT/Qwen3-Coder-Next-FP8; QWEN_FP8_DIR may be set only to that exact path. Broad roots, temporary roots, symlinks, hard-linked files, untrusted ownership/modes, dirty pack trees, and alternate origins fail closed.

The compatibility commands ./scripts/serve_spark.sh and ./scripts/serve_spark_vllm.sh invoke the same FP8/vLLM launcher. It is not a GGUF or llama.cpp path.

Check the local endpoint:

curl --fail --silent \
  -H "Authorization: Bearer $VLLM_API_KEY" \
  http://127.0.0.1:8001/v1/models

OPENAI_BASE_URL=http://127.0.0.1:8001/v1 \
OPENAI_API_KEY="$VLLM_API_KEY" \
SMOKE_MODEL=local-qwen3-coder-next \
/usr/bin/python3 -I -S eval/agent_smoke/run_smoke.py

vLLM applies VLLM_API_KEY to requests whose paths begin with /v1, /v2, or /inference. Other paths are not authenticated by that mechanism. This Preview profile is therefore for a trusted single-user host: Docker host-loopback publication reduces network reachability, but it is not isolation from other local processes/users, a firewall, or whole-server authentication. The key is also passed through Docker's environment and is inspectable by local root or Docker administrators; Docker environment variables are not a secret store.

Remote access

Direct LAN binding is deliberately disabled. Use an authenticated SSH tunnel:

ssh -L 18001:127.0.0.1:8001 spark

Then connect the client to http://127.0.0.1:18001/v1 with the same API key. For a shared service, a separately reviewed TLS proxy must authenticate and authorize the entire vLLM service, including paths outside /v1, /v2, and /inference; do not change HOST to 0.0.0.0.

Exact provenance and runtime pins

Field	Value
Upstream base	Qwen/Qwen3-Coder-Next
FP8 repository	Qwen/Qwen3-Coder-Next-FP8
FP8 revision	da6e2ed27304dd39abadd9c82ef50e8de67bdd4c
Expected shards	exactly 40
Verified snapshot files	exactly 52, including tokenizer/template/parser support files
Manifest SHA-256	d7269c078ec614ae57792ec63da00b3e13cf88e61f421ef4bcf4bdaac54b0190
vLLM version	0.25.1
vLLM commit	752a3a504485790a2e8491cacbb35c137339ad34
OCI index digest / image reference	sha256:e4f88a835143cd22aee2397a26ec6bb80b3a4a6fe0c882bcbc63822904766089
Selected platform	linux/arm64
arm64 platform-manifest digest	sha256:2cc49b81319f7a66a33dd8bd63a7bfddae079122b33ce51989b6828a1f038c37
Image-config digest / Docker image ID	sha256:30a38a1d74a17365eca400e83ffd885b250e0c8c0d3c5b508afa8c412d2ddf95
Docker boundary	clean client environment · explicit local Unix socket · exact image · remote daemon variables removed
Pack source	canonical GitHub origin + externally supplied 40-hex commit + clean tracked tree

scripts/pull_official_fp8.sh writes both .verified-revision and an atomic PULL_COMPLETE marker containing the repository, revision, checksum-manifest digest, verified file count, and verified shard count. The launcher validates all of those fields before loading the model.

Evidence: what 2/2 means

The historical receipt in results/agent_smoke.json records two successful API/chat availability prompts: an exact pong response and a small Python add response. Neither prompt supplied tools. Therefore 2/2 is not evidence of tool-call formatting, routing, tool execution, closed-loop recovery, coding benchmark quality, throughput, concurrency, or long-horizon reliability.

That historical probe used a floating vLLM image and does not validate the newly pinned runtime. It is retained as historical evidence, not promoted as a launch headline. A Preview listing does not imply SAQS certification.

For optional post-listing protocol work, the qwen_native_closed_loop harness executes three dynamic tool roundtrips, including dependency chaining and error recovery. It has no checked-in pass receipt and is not a new prerequisite for the Preview listing.

Known limits

• One active request is the day-0 default; concurrency is unmeasured.
• The tested public ceiling is not established beyond the historical 8K probe; 32K is a conservative launcher ceiling, not a measured effective-context claim.
• The pack has no reference-vs-quant retention study and hosts no weights.
• The native parser is configured, but tool correctness remains unmeasured.
• GGUF helper files are retained for research history only. They are not the supported day-0 path and scripts/serve_spark.sh does not serve GGUF.
• Trust-remote-code is not enabled.
• VLLM_API_KEY covers only the /v1, /v2, and /inference path prefixes. Other paths are not authenticated by that mechanism, and host loopback is reachable by local software/users. Use a whole-service boundary when shared.
• Path, ownership, link, Docker-context, image, residency, and port checks are point-in-time guards. The complete 52-file snapshot and Docker boundary are revalidated after both locks. Same-UID and privileged mutation remain outside this script’s sandbox.
• Pull and serve share a Qwen operation lock; serve also uses the same host-wide model-residency lock as the Laguna launcher. Stale locks require manual inspection.

Troubleshooting

Failure	Meaning / action
checksum-manifest mismatch	Use the pack’s unchanged results/SHA256SUMS.fp8
missing or invalid marker	Re-run scripts/pull_official_fp8.sh
pack commit/origin/dirty-tree rejection	Use the canonical clean checkout at the independently recorded release commit
Docker context/socket rejection	Restore local default context on the root-owned Unix socket; remote daemons are unsupported
image ID/platform rejection	Pull the exact linux/arm64 digest shown above; do not substitute a tag
API-key rejection	Set a random URL-safe VLLM_API_KEY of 32–256 characters
stale operation/residency lock	Inspect the host and lock owner; never auto-delete an unexplained lock
resident process/container	Stop every sibling model server, then retry
available memory below 96 GiB	Restore the required single-model state; do not lower the bar casually
swap already used	Stop memory pressure, clear swap safely, and retry
non-loopback host rejected	Use SSH tunneling or an authenticated TLS proxy
model id error	Request local-qwen3-coder-next

Release integrity

• Source is freeze-attested (2026-08-02); class remains Preview only. Exact-runtime DGX Spark evidence is still pending — not runnable_day0 / not hero.
• LAUNCH_GATE.md separates Preview listing readiness from future measured claims.
• SPARK_AGENTIC_QUANT_STANDARD.md defines the shared future certification bar.
• results/RELEASE_MANIFEST.json binds the public class, exact runtime, and conditional promotion posture.
• smoke != headline; verification != gate clearance.

License and affiliation

Pack code and documentation are Apache-2.0; see LICENSE. Upstream weights remain governed by their upstream repository. This is an independent personal deployment pack, not an Alibaba, Qwen, NVIDIA, or vLLM product.