Downloads · 30 days
65
28% of all-time downloads
0xSero/Laguna-S-2.1-Hybrid-2.75bpw
Laguna-S-2.1-Hybrid-2.75bpw is a text generation model from 0xSero. Use it when you need the model to write or continue text. It is set up for vllm. The card lists the license as openmdw-1.1.
Verified public package. The weight transform, TP2 runtime, parser, structured-output, 195k-context, throughput, package-manifest, and Hub-object verification gates completed before public release.
Downloads · 30 days
65
28% of all-time downloads
All-time downloads
230
Public
Repo size
50.1 GB
Likes
1
Public
Click a slice to open those files.
.safetensors12 GB · 100%
From the Hugging Face model README
Verified public package. The weight transform, TP2 runtime, parser, structured-output, 195k-context, throughput, package-manifest, and Hub-object verification gates completed before public release.
Laguna S 2.1 Hybrid 2.75bpw is a two-tier, expert-level quantization of
poolside/Laguna-S-2.1-NVFP4.
It preserves every expert and the source router topology. REAP saliency is used
only to choose the NVFP4 hot tier; it is not expert pruning.
Each of the 47 sparse MoE layers has 256 routed experts:
| Tier | Experts per sparse layer | Storage |
|---|---|---|
| Hot | 96 | source NVFP4, retained byte-exact |
| Tail | 160 | calibrated EXL3 Trellis-2 |
This is a nominal 2.75-bit mixture across the expert weights. Shared experts, the router, attention, embeddings, LM head, and other non-expert tensors remain in their source representation. All 256 experts remain routable.
The packaged source shards are compacted: they contain only the 96 selected NVFP4 routed experts per sparse layer. The other 160 routed experts are present only in the calibrated Trellis-2 tail files, so the package does not duplicate their original NVFP4 tensors. The release manifest seals 50,111,846,955 bytes (approximately 46.67 GiB), excluding the manifest file itself. The 2.75-bpw label describes routed-expert weight storage, not the whole repository after non-expert tensors and metadata.
The hot experts are ranked independently in each sparse layer using router- weighted expert-output activation saliency over a held-out calibration mixture covering coding, tool and agentic calls, reasoning, mathematics, science, terminal work, CUDA, cybersecurity, and long-context prompts. The score is the mean, over tokens routed to an expert, of the applied router weight multiplied by that expert's output L2 norm. The release includes the exact tier map, source revision, calibration provenance, and artifact hashes.
The hybrid requires the accompanying EXL3-tail runtime and is not a drop-in stock-vLLM checkpoint. The runtime supports single-Spark TP1 by reconstructing both stored tail slices on one device, and two-Spark TP2 by loading one stored slice per rank. The package includes a reproducible Dockerfile, the pinned EXL3 source archive and ARM patch, the runtime overlay, and the exact launch recipes.
Anemll/dspark-vllm-gx10ghcr.io/anemll/dspark-vllm-gx10:0.1.1runtime/DockerfileThe sealed validation bundle records:
The source checkpoint has no native MTP tensors. The pinned Laguna DFlash draft
model was tested but rejected for this compressed hybrid after preserved
sustained-decode divergence/nontermination evidence. The verified default is
therefore LAGUNA_SPECULATIVE_MODE=none; DFlash remains an explicit
experimental opt-in and is not claimed as validated.
No Terminal-Bench 2.1 quality score is claimed in this card.
poolside/Laguna-S-2.1-NVFP4This derivative remains subject to the upstream OpenMDW-1.1 license and should be used consistently with the upstream model's acceptable-use guidance.