Image-Text-to-Text
Safetensors

You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

By accessing Ultramarine 1 Flash, you agree to the Ultramarine 1 Flash License. Model Providers, as defined in the license, must enter into a separate agreement with Paper Instruments.

Log in or Sign Up to review the conditions and access this model content.

Ultramarine 1 Flash LoRA

Paper Instruments

Introduction

This repository contains the unmerged LoRA for Ultramarine 1 Flash. Apply it to GLM-5.3-Flash to serve Ultramarine alongside GLM-5.3-Flash traffic on the same base weights. For dedicated deployment, use the merged weights in paperinstruments/Ultramarine-1-Flash.

Read the full technical report.

Adapter

Setting Value
Base model zai-org/GLM-5.3-Flash
Rank 64
LoRA alpha 64 (scaling 1)
Factor precision BF16
Coverage KDA and sparse-attention projections, dense MLPs, shared experts and routed experts
Files adapter_model.safetensors (10.2 GB), adapter_config.json

The adapter keeps the shared-outer layout it was trained in, so it needs the loader support below. Keep adapter_config.json next to adapter_model.safetensors.

Runtime Requirements

The files follow PEFT naming, but the shared-outer layout needs model-specific loader support. As of October 28, 2026, stock PEFT and released versions of vLLM and SGLang do not load it correctly.

  • vLLM: requires #56327 (GLM projection mapping) and #56718 (kv_b_proj LoRA during MLA prefill), both unmerged. Serve with BF16 LoRA, enable_moe_shared_loras=true and max_lora_rank=64.
  • SGLang: requires #44862 for KDA projection support. It is merged into the sglang-miles branch and included in Miles v0.1.3 (radixark/miles:v0.1.3), but not in an SGLang release.

We have verified loading and generation on both engines with BF16 KV cache, but not numerical parity with the merged weights. With speculative decoding, the MTP draft runs without the adapter. We have not tested this MTP setup.

For sampling and harness recommendations, see Ultramarine 1 Flash.

License

Ultramarine 1 Flash is released under the Ultramarine 1 Flash License.

Citation

@techreport{paperinstruments2026ultramarine,
  title       = {Ultramarine 1 Flash: Scaling Non-Verifiable RL},
  author      = {{Paper Instruments}},
  institution = {Paper Instruments},
  year        = {2026},
  url         = {https://www.paperinstruments.com/blog/ultramarine}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for paperinstruments/Ultramarine-1-Flash-LoRA

Adapter
(9)
this model