Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

5.2 KB
Raw Blame

DeepSeek V4 Flash Vision and Laguna XS 2.1 Vision

This directory is the reproducibility record for two WebBrain experiments that add a frozen MoonViT image tower and a trained patch-merger projector to frozen mixture-of-experts language models.

The record was assembled on 2026-08-04. It distinguishes observed results from planned experiments and pins every published model repository to the commit that was inspected. It is a technical report, not a peer-reviewed paper or a claim of benchmark parity with an upstream multimodal model.

Document map

  • Research report: motivation, method, measured results, comparative analysis, limitations, and conclusions.
  • Reproducibility runbook: exact revisions, commands, checkpoint policy, verification gates, publication layout, and recovery notes.
  • Source appendix: a curated code snapshot, archived run metadata, published model cards/manifests, and checksum instructions. No model weights or dataset images are duplicated in this Git repository.

Final model repositories

Model package Role Inspected revision
DeepSeek V4 Flash Vision NVFP4 Complete pinned NVFP4 text package plus MoonViT tower, projector, and custom SGLang source 12b653e63329ac3c20395f9aeeb1bb8264d2db8b
DeepSeek V4 Flash Vision BF16 BF16 vision source overlay; deliberately not a full 291B BF16 text checkpoint 60c441aa1c7386387c89ddaff703136395ad8d8b
DeepSeek V4 Flash Vision Training Archive Private bring-up, calibration, recovery, and code archive; it predates the completed 100K run Private, mutable access-controlled repository
Laguna XS 2.1 Vision NVFP4 Complete pinned NVFP4 text package plus MoonViT tower and projector ce108f0f3764a18a1f5f7d14ecefa90485ea6e52
Laguna XS 2.1 Vision BF16 Complete pinned BF16 text package plus MoonViT tower and projector b06063d00b73f7713ef1c28f8247eed488c73faf
Laguna XS 2.1 Vision Projector 100K Private full projector/optimizer checkpoint history and run evidence 4f5d6359867fa55315a2bec4b568901bfdeca5e5

What is complete

  • Both final projectors were trained on 100,000 examples for one epoch and published with their frozen BF16 MoonViT towers.
  • The DeepSeek adapter preserves the discrete token IDs required by DeepSeek V4's hash-routed experts while separately injecting projected image embeddings.
  • DeepSeek's pinned SGLang/Blackwell profile passed full NVFP4 loading and two consecutive end-to-end image requests on NVIDIA B200 with four-way tensor parallelism.
  • The Laguna private archive retains the full 782-step projector and optimizer checkpoint history, its training log, final state, hashes, and two code snapshots.
  • Public NVFP4 and BF16 package layouts, source model revisions, component sizes, and component SHA-256 values are recorded in the runbook and artifact manifest.

What remains incomplete

  • No controlled MoonViT-versus-Qwen tower A/B result was completed. The Qwen arm in the source tree is an experimental plan, not a reported result.
  • No standard vision benchmark suite is archived for either final package, so this report does not make quality claims beyond finite training, loss behavior, component integrity, and the explicitly listed inference gates.
  • Laguna's main BF16 and NVFP4 revisions are not end-to-end image endpoints. An unvalidated integration candidate exists only on the NVFP4 sglang-integration branch at 740a6ccc05441a7ea0426c1d2536d5cd031adb86; no loader or image smoke result is recorded.
  • DeepSeek's custom SGLang package still requires its pinned source patch and is not stock-SGLang compatible. The B200 result is a bounded functional smoke test, not text-only parity, a broad quality benchmark, or production validation.
  • The DeepSeek private archive does not contain a complete final 100K checkpoint history; the public final projector is reproducible as an artifact, but the full run telemetry is not as complete as Laguna's.

Reproduction rule

Use commit-pinned URLs and verify hashes before loading any component. A mutable model page URL is included for convenience, but the commit URL and SHA-256 are the reproducibility identity. Never infer serving readiness from the presence of model weights alone.