You can deploy this model and start earning money today!

cyjin-yl/
DeepSeek-V4-Flash-0731-Vision

Views

No views yet

[!WARNING] ARCHIVED — this page is a historical research record, not a usable DeepSeek VLM.

The V100/32GB experiments in the companion repository did not produce an external-MoonViT projector that passed the preregistered causal capability gate: the correct image was not significantly better than both the blind and shuffled-image conditions, and ScreenSpot grounding did not improve. We therefore stopped this research line on 2026-08-08. The project was constrained by the inability to rent Blackwell-class hardware within the fixed budget; no paid server or full DeepSeek-V4-Flash-0731 training was run.

This repository contains MoonViT/control artifacts, not a validated DeepSeek-V4-Flash-0731 vision checkpoint. It does not contain the full DeepSeek weights, a verified FP4/FP8 image-to-answer run, or a deployable image-generation service. Do not read the Qwen/SmolLM controls as evidence that DeepSeek has acquired vision.

Full audit and the consolidated record of positive, negative, and failed experiments: FINAL_ARCHIVE_REPORT.md. The source repository is archived: cyjin-yl/moonvit-deepseek-v4-glue.

For comparison only, webbrain-one/DeepSeek-V4-Flash-0731-Vision-NVFP4 publishes a separate external overlay. Its public manifest says gpu_validated_for_this_0731_package=false and does not publish the paired ScreenSpot/TextVQA/DocVQA/OCRBench evidence used here. This is an evidence-boundary note, not a claim that the project is fraudulent.

Archived DeepSeek-V4-Flash-0731 Vision Projector research artifacts

Final status (2026-08-08)

Question	Result
Can the MoonViT feature/cache and projector software run?	Yes, on small frozen receivers and the tiny DeepSeek software fixture.
Did an external MoonViT projector pass the fixed visual-causality gate?	No. No arm achieved `vision > blind` and `vision > shuffled` with a ScreenSpot click gain.
Strongest external-projector result	Qwen2.5-7B + exact K3/MoonViT-V2, 57,600 examples; training was numerically healthy but ScreenSpot vision click (3.30%) was not above blind (3.46%).
Independent positive control	Native Qwen3.5 VLM (its own visual tower); this is not a MoonViT-projector result.
DeepSeek-V4-Flash-0731 Gate D	NO-GO: full-weight placement, FP4/FP8 backward, checkpoint round-trip, and real image generation were not validated.

The locally generated projector/LoRA/health checkpoint payloads were uploaded under archive/checkpoints/ in commit a40630b1 before their workstation copies were removed. The archive now contains 84 files (1,302,790,146 bytes), including 14 late 7B capacity-control files under archive/checkpoints/late_capacity_controls/; paths and sizes were checked with HfApi. The lightweight manifests/configs remain in the archived Git repository.

The historical files below are retained so that the negative result is reproducible. They are not a release claim.

Historical page content (preserved)

Goal of this repo: a trained projector that connects a frozen MoonViT-V2 (Kimi K3 vision tower, extracted to vision_tower_k3/) to a frozen DeepSeek-V4-Flash-0731 language model, following the community GLM-5.2-with-vision recipe (Harry Partridge / 0xSero reproduction). Code: https://github.com/cyjin-yl/moonvit-deepseek-v4-glue

Directory map — read this before using any weights

Path	What it is	Backbone
`vision_tower_k3/`	Extracted MoonViT-V2 tower weights + configs + sha256 MANIFEST. Frozen, backbone-independent.	(vision tower)
`eval/`	Benchmark reports per run tag (per-record predictions + metadata + SUMMARY). Public raw outputs.	—
`gate_b_qwen05_v100/`	LOCAL CONTROL experiment — NOT DeepSeek weights. Projector trained against frozen Qwen2.5-0.5B-Instruct on a single V100 (2,000 steps, full 59k mix). Validates the train/eval/upload pipeline and provides the small-backbone control group.	Qwen2.5-0.5B
`gate_b_smoke_smollm135_v100/`	Earlier pipeline dress rehearsal — NOT DeepSeek weights. 400-step smoke run against frozen SmolLM2-135M-Instruct.	SmolLM2-135M

As of 2026-08-04 this repo contains NO DeepSeek-V4-derived projector. Every projector*.safetensors here belongs to a small-backbone control run and only fits its respective backbone hidden size (896 / 576). The DeepSeek projector (4096 hidden) will land under its own clearly-named directory after the rental training run.

How a projector is meant to be used

load_deepseek_flash_0731(projector_dir) in the glue repo loads the frozen 0731 backbone, the frozen MoonViT-V2 tower and a projector directory (projector_config.json + projector.safetensors) as three separately auditable weight sources. Inference-side integration (vLLM/SGLang patch surface) is documented in docs/inference-integration.md of the glue repo.

Provenance

Vision tower: extracted from Kimi K3 shard model-00096-of-000096.safetensors (sha256 recorded in vision_tower_k3/MANIFEST.json); the full K3 repo is never needed downstream.
Train/eval data: cyjin-yl/moonvit-dsv4-data (sources, revisions, sha256 and decontamination report included).
Eval slices: selection half (--record-slice even) of each benchmark; results always paired with the blind (no-image) baseline and, for the Gate B run, a random-projector control.

cyjin-yl/DeepSeek-V4-Flash-0731-Vision