Modelpublic

Baekpica/Solar-Open2-250B-GGUF

sourceHugging Faceotherupdated 27d agoView on Hugging Face

0likes361downloads

Model Card

Solar Open2 250B GGUF

Built with Solar.

Full-topology GGUF conversions of `upstage/Solar-Open2-250B`, preserving all 48 layers, 320 routed experts, and one shared expert. This is an independent conversion and is not an official Upstage release.

Support my work

I work on making large language models practical on hardware they were never really designed to fit on — through mixed quantization, inference optimization, custom kernels, and serving experiments.

While much of the development happens on local hardware, calibration, profiling, and large-scale validation often require expensive on-demand GPUs.

Contributions help pay for that compute, storage, and testing infrastructure, so I can keep experimenting and publishing the results openly.

Variants

Variant	Split	Purpose
BF16	11 shards, 500,718,700,672 bytes (466.331 GiB)	High-precision converted reference
Q8_0	11 shards, 266,136,847,072 bytes (247.859 GiB)	Full-model imatrix and quality reference

The exact per-shard hashes are in BF16-SHA256SUMS and Q8_0-SHA256SUMS. Each shard is uploaded as an independent resumable Hub commit; the two variants use disjoint filenames and can be downloaded independently.

Provenance

—Source revision: upstage/Solar-Open2-250B@9190fbe63a2ad8e17fc766ccceb36de7c66f004b
—Source tensor count: 47,019
—Source bytes: 500,617,750,484
—GGUF architecture: solar-open2
—Context metadata: 1,048,576 tokens, NoPE
—Converter base: ggml-org/llama.cpp@6ea215d171fd31df943bf1ac8227129f2b963160
—Solar conversion/runtime patch SHA256: 7c267850de7a08af66a9bd64e9d4a3b33e085a4d2888d94a4c3d919f87283e44

The BF16 conversion uses --outtype bf16 --split-max-size 48G. The Q80 conversion uses `llama-quantize`, `--pure`, `--keep-split`, Q80 token embeddings, and a Q8_0 output tensor.

Validation status

Both complete split sets passed GGUF metadata, split ordering, tensor count, topology, bounds, and exact tensor-type checks with zero errors. Each contains 1,083 GGUF tensors: BF16 has 674 BF16 + 409 F32 tensors, and Q80 has 674 Q80 + 409 F32 tensors.

The Q80 model was loaded resident across four H100 NVL GPUs and completed a 32-prompt greedy reference run plus no-think quality sanity prompts covering Korean, English, Japanese, code, structured JSON, reasoning, and long-text summarization. The structured-output sample parsed as valid JSON and no NaN, malformed-output, or broken-generation regression was observed. The four-H100 run is development evidence only, not a DGX Spark `sm121` serving or performance claim.

License and attribution

The source model and these conversions are distributed under the included Upstage Solar License. The model name begins with Solar, and this model card prominently states “Built with Solar” as required by that license.