DeepSeek-V4-Flash-REAP25-Spark-Agentic

Parameters

Total	284B (official DeepSeek card)
Active / token	13B
Context	1M tokens
Arch	MoE · DeepseekV4ForCausalLM
Default OS path	Peer REAP25 + Pulsar path ~85 GiB experimental (weights not in this repo)
Source	deepseek-ai/DeepSeek-V4-Flash
This pack	docs / evaluation scaffold only — no weights · unmeasured · not runnable

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: deepseek-ai/DeepSeek-V4-Flash for findability only — this is not a finetune/adapter/quant of that checkpoint.

DOCS-ONLY HOLD · NO_HERO · NOT RUNNABLE · NO WEIGHTS · UNMEASURED

This personal repository documents a possible single-DGX-Spark investigation using a third-party REAP25 derivative. It is not a model checkpoint, quantized model, finetune, adapter, runtime distribution, benchmark result, or DeepSeek release. It contains no weight files and performs no download or inference.

Current release lock

Field	Locked value
Release state	docs_only_hold
Installable / runnable	false / false
Measured	false
Hero CTA	null (NO_HERO)
Weights in this repo	none
Defined text checks	2, unexecuted
Tool-call cases / measurements	0 / 0
Download entrypoint	disabled; exits 2
Serve entrypoint	disabled; exits 2
Client and smoke entrypoints	disabled; exit 2

The canonical machine-readable state is config.yaml. The release lock is results/launch_lock.json.

Reference facts — not pack measurements

Field	Reference-only value
Upstream architecture	DeepSeek advertises 284B total / 13B active MoE
Upstream context	DeepSeek advertises 1M tokens
Third-party TYPE40 file	91,321,404,640 bytes (85.049686 GiB), absent and unverified here
This repository	docs only; zero weights; zero runtime binaries; zero model executions
DGX Spark claim	no fit, usable-context, quality, speed, or stability claim

The upstream facts describe deepseek-ai/DeepSeek-V4-Flash, not this pack. The TYPE40 size describes the pinned third-party file, not proven runtime residency.

Immutable artifact reference

The only candidate pinned by this scaffold is a third-party artifact. It is not downloaded, mirrored, endorsed, or verified locally by this pack.

Field	Value
Repository	twaggs88/DeepSeek-V4-Flash-REAP25-DSpark-ds4-GGUF
Revision	4b853e8b0adc6ac7c4499faf8de1ca439b116822
File	ds4flash-v5mx-reap25-type40-mxfp8lt-dspark-v1.gguf
Size	91,321,404,640 bytes (85.049686 GiB)
SHA-256	000974720296f2cad17ac0525796f4bb9ceaac9f4015ed61af3fba445dfb1039
Attribution	third-party REAP25 / ds4 research artifact

DSpark in the artifact name refers to a DeepSeek speculative-decoding module; it does not mean NVIDIA DGX Spark. A file size below 128 GB also does not prove runtime residency, usable context, stability, or performance on DGX Spark.

Runtime compatibility

The artifact publisher documents custom tensor formats accepted only by the ds4 engine and explicitly says that common GGUF loaders do not load it. The previous Laguna llama-server route in this pack was therefore incorrect and has been removed. Do not use this TYPE40 file with llama.cpp, Ollama, vLLM, Transformers, or a Laguna server binary.

The future reference runtime is:

Field	Reference-only pin
Repository	tylerwagler/pulsar
Tag	v0.3.1
Commit	536466ccf7f59b031c0c9e4231c39ebfb591684e
Binary	ds4-server
CUTLASS submodule	e05f953a5b3d38adc240df2ff928e0421c2abba3
Pack status	not vendored, built, installed, or tested

This pin is documentation, not an enablement path. No command, flag, or environment variable in this repository starts a server.

Why every entrypoint fails closed

The earlier implementation could float repository metadata, fall back to an unsafe size floor, and unlock an incompatible server through environment variables. The current scripts have no network, filesystem-staging, client, or server code. Invoke them explicitly (downloaded Hub files may not preserve executable mode):

./scripts/pull_peer_gguf.sh
./scripts/serve_spark.sh
/usr/bin/python3 -I -S eval/agent_smoke/run_smoke.py
/usr/bin/python3 -I -S hermes/sample_client.py

Every command exits 2. The shell stubs also print the immutable references; the Python stubs print only their disabled status.

See docs/WEIGHTS_LOCAL.md for the verification and measurement work required before a future, separately reviewed runnable release.

Evaluation honesty

eval/agent_smoke/cases.json contains two simple text-response definitions. They were not run, contain no tool schemas, and cannot support an agentic, function-calling, safety, coding, long-context, throughput, or reliability claim. There is no receipt.

A future agentic evaluation must at minimum bind the exact artifact and runtime pins, supply real tool schemas, execute the tools in a sandbox, validate native structured calls, publish raw outcomes, measure memory and stability, and keep evaluation tasks out of any calibration data. The shared requirements are in SPARK_AGENTIC_QUANT_STANDARD.md.

August 3 scope

August 3, 2026 is a documentation co-list target only. It is not a model launch, runtime launch, benchmark launch, weight release, or hero campaign. The CTA stays null for this release. See LAUNCH_AUG3.md.

This pack cannot honestly generate model-file download counts while it hosts no model file. Views, likes, clones, and downstream downloads must not be described as checkpoint downloads. A future checkpoint repository should be published only with redistribution permission, immutable files, checksums, a compatible runtime, reproducible tests, and independently verifiable receipts.

Identity and attribution

• Upstream model reference: deepseek-ai/DeepSeek-V4-Flash
• Third-party artifact: twaggs88/DeepSeek-V4-Flash-REAP25-DSpark-ds4-GGUF
• This pack: hizrianraz/DeepSeek-V4-Flash-REAP25-Spark-Agentic
• Legacy identifier: hizrianraz/DeepSeek-V4-Flash-Spark-Agentic (redirect only)

This is an independent personal project. It is not affiliated with or endorsed by DeepSeek, the artifact publisher, NVIDIA, or Ainfera. diy_gguf: false.

Repository map

Path	Purpose
config.yaml	canonical docs-only state and immutable pins
INSTALL.yaml	non-installable pointer to config.yaml
docs/WEIGHTS_LOCAL.md	artifact verification requirements
hermes/	disabled compatibility notes
eval/agent_smoke/	two unexecuted text definitions
tests/	fail-closed docs-only contract checks; no model execution
results/	locks; no benchmark receipts

Updated: 2026-07-30 — audited docs-only NO_HERO scaffold.