---

base_model: upstage/Solar-Open2-250B

base_model_relation: quantized

license: other

license_name: upstage-solar-license

license_link: LICENSE

language: [en, ko, ja]

pipeline_tag: text-generation

library_name: gguf

tags: [gguf, solar, moe, kda, mixed-quantization, dgx-spark]

---

Solar Open2 250B — Mixed-Quant GGUF

Built with Solar.

**A structurally intact Solar Open2 250B-class MoE compressed to 88.97 GiB

and resident on one 128 GB DGX Spark. The 1,048,576-token release gate remains

open.**

This is not a distillation, pruned model, expert-dropped variant, or

layer-truncated conversion. All 48 layers, all 320 routed experts per layer,

the shared expert, and the original [GQA, KDA, KDA, KDA] x 12 schedule are

preserved. The final artifact contains 1,083 tensors. Precision is assigned

by what each tensor does rather than by one uniform bit budget.

The complete GGUF is 88.973 GiB across 11 shards. It has now been loaded as a

native sm_121a CUDA-KDA family on one GB10 with aligned IQ2/Q8 expert

artifacts, run through a full-model 2K BF16/FP8/hybrid/FP4 KV comparison at

up to 328 tok/s prefill, passed the staged 2K-64K KDA lifecycle regression

(snapshot, replay, fork, cancellation), and completed a measured 131,072-token

prefill and decode with every needle retrieval passing.

Full-depth 512K/1M retrieval semantics and 1M API serving are still work in

progress; this card does not claim a completed one-Spark 1M release.

| | |

|---|---:|

| Architecture | Solar Open2, 48 layers, [GQA, KDA, KDA, KDA] x 12 |

| Routed experts retained | 320 / 320 per layer |

| Tensors | 1,083 |

| Artifact | 88.973 GiB, 11 shards |

| Approximate weight BPW | 3.05 |

| Resident model copy on GB10 | 88.96 GiB, measured |

| Native CUDA target | GB10 / sm_121a, validated |

| Source context limit | 1,048,576 tokens |

| Full 1M execution | not yet validated |

This is an independent conversion of

upstage/Solar-Open2-250B,

not an official Upstage release.

Artifact

| Variant | Split | Exact size | Approximate BPW |

|---|---:|---:|---:|

| MXQ-v1 | 11 shards | 95,533,532,160 bytes (88.973 GiB) | 3.05 |

Start with

Solar-Open2-250B-MXQ-v1-00001-of-00011.gguf; compatible split-aware runtimes

load the remaining ten shards automatically. Exact per-shard hashes are in

MXQ-v1-SHA256SUMS.

The weight files are pinned to Hub revision

cd504d0ce46c0850dd34c319659e020a17ea7194. Later model-card commits change

README.md only; they do not change any shard byte or hash.

hf download Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF \
  --revision cd504d0ce46c0850dd34c319659e020a17ea7194 \
  --include 'Solar-Open2-250B-MXQ-v1-*.gguf' \
  --include MXQ-v1-SHA256SUMS \
  --local-dir ./Solar-Open2-250B-MXQ-v1

The full BF16 and full-model Q8_0 reference conversions are published

separately in

Baekpica/Solar-Open2-250B-GGUF.

Precision recipe

Quantization is assigned by module role, keeping numerically sensitive and

always-active paths conservative while placing the compression burden on the

routed experts.

| Tensor group | Type | Why |

|---|---|---|

| Token embedding, LM head | Q8_0 | token and logit fidelity |

| GQA Q/K/V/O/g projections | Q8_0 | long-context stability |

| KDA matrix projections | Q8_0 | recurrent-state stability |

| Shared expert gate/up/down | Q8_0 | every token uses it |

| Routed gate/up, layers 4-43 | IQ2_XXS + imatrix | dominant parameter mass |

| Routed down, layers 4-43 | Q3_K + imatrix | more conservative accumulation path |

| Routed experts, layers 0-3 and 44-47 | Q4_K + imatrix | protect the edge blocks |

| Router, norms, KDA controls/norms/convs | F32 | small and decision-sensitive |

The final type inventory is **409 F32, 530 Q8_0, 80 IQ2_XXS, 40 Q3_K, and

24 Q4_K tensors**. quant-recipe-v1.yaml, the exact tensor-type assignment,

and verify-v1.json are published with the artifact.

Calibration

The Solar chat-template calibration mix follows the published distribution

from Baekpica/Solar-Open2-120B-A15B-REAM-148E-Healing-Mix: instruction chat

22%, Cascade stage-1 reasoning 16%, Cascade stage-2 reasoning 16%, Korean 16%,

other multilingual 12%, finance 6%, SWE 6%, and algorithmic code 6%.

The importance matrix processed 590 x 512-token chunks. Layers 1-47 reached

strict 320/320 routed-expert coverage. Layer 0 retains one explicit, pinned

exception of 29 zero-activation experts; their IDs and the acceptance gate are

recorded in verify-imatrix.json. No precision was lowered to hide that

exception.

Provenance

| | |

|---|---|

| Source model | upstage/Solar-Open2-250B@9190fbe63a2ad8e17fc766ccceb36de7c66f004b |

| Official implementation oracle | upstageAI/transformers@ef5c37485091057070b36dd0620798f74988eb01 |

| llama.cpp conversion base | ggml-org/llama.cpp@6ea215d171fd31df943bf1ac8227129f2b963160 |

| Fixed weight revision | cd504d0ce46c0850dd34c319659e020a17ea7194 |

| Reproduction repository | Baekpica/solar-open2-mixed-ds4 |

| Spark serving engine | Baekpica/ds4 |

| Measured engine code | c36e3f2 |

The Solar conversion/runtime patch SHA-256 from the H100 handoff is

7c267850de7a08af66a9bd64e9d4a3b33e085a4d2888d94a4c3d919f87283e44.

Model structure

Solar Open2 alternates one NoPE GQA layer with three KDA recurrent layers,

repeated 12 times. Each layer retains 320 routed experts plus one shared

expert. This combination matters for memory accounting:

• only 12 layers grow a conventional attention KV history;
• 36 KDA layers keep fixed recurrent and convolution state;
• the 320-expert routed stack contains most of the stored parameters;
• runtime KV compression changes context memory, not the published GGUF

weights.

Runtime compatibility

Serving solar-open2 requires a runtime that implements its GQA/KDA family.

The measured Spark path is

Baekpica/ds4:feature/solar-open2-model-loader

at 197a8b9, built with make cuda-spark. Stock runtimes may not yet recognize

the architecture. Performance reported below applies only to that pinned ds4

build.

Verified DGX Spark performance

Measured on NVIDIA GB10 (sm_121), 121.6 GiB unified memory, driver 595.71.05,

CUDA 13.3, and Linux 6.17.0-1029-nvidia.

Startup builds 381 aligned CUDA artifacts (39.60 GiB device, ~100 s) that

replace the raw IQ2_XXS gate/up expert reads and add aligned Q8 dense copies,

then promotes the 47.95 GiB Q3_K/Q4_K expert remainder into CUDA-owned

memory. The engine runs the routed IQ2 experts on a dequant-to-registers

pair kernel, KDA prefill as 64-token chunked delta-rule UT transforms, and

compressed-KV attention on HMMA tiles.

One resident engine evaluated the same deterministic 2,048-token prefix with

four runtime KV layouts. Decode is the latency of one identical continuation

token selected from the BF16 prefill.

| Runtime KV | Prefill time | Prefill rate | One-token decode | Prefill top-1 | Decode top-1 |

|---|---:|---:|---:|---:|---:|

| BF16 | 15.638 s | 130.962 tok/s | 138.362 ms | 11047 | 27294 |

| FP8 | 6.245 s | 327.958 tok/s | 96.950 ms | 11047 | 27294 |

| K-FP8 / V-FP4 hybrid | 6.242 s | 328.087 tok/s | 95.629 ms | 11047 | 27294 |

| FP4 | 6.256 s | 327.343 tok/s | 95.628 ms | 11047 | 27294 |

All four layouts produced finite logits and the same prefill and decode top-1

in this measured run (the BF16 leg runs first and carries the one-time lazy

cache warmup). The staged KDA lifecycle regression then passed on this exact

configuration with bit-stable replay at every stage:

| Lifecycle stage | Extension prefill | Rate | Snapshot |

|---:|---:|---:|---:|

| 2,048 | 11.95 s | 171.3 tok/s | 207 MiB / 0.13 s |

| 8,192 | 17.76 s | 345.9 tok/s | 353 MiB / 0.22 s |

| 32,768 | 90.91 s | 270.3 tok/s | 938 MiB / 0.52 s |

| 65,536 | 172.31 s | 190.2 tok/s | 1.68 GiB / 1.05 s |

At 131,072 tokens (130,800-token rendered needle prompt, hybrid KV) the

measured cold prefill was 819.7 s (159.6 tok/s) with 4.13 tok/s greedy

decode and 3.709 GiB of context pools, completing with no OOM, no

non-finite logits, and all four needle retrievals (begin, middle, end,

latest version record) correct. Rates fall with depth because the 12 full-attention GQA layers pay

the quadratic term. These numbers are not extrapolated beyond what is

listed.

Validation status

• All 11 shards passed ordering, bounds, topology, tensor-count, tensor-family,

and exact tensor-type verification.

• Four-H100 development validation completed deterministic Korean, English,

Japanese, code, arithmetic, JSON, and constrained-list checks without NaN or

malformed structured output.

• One GB10 completed native sm_121a loading, the aligned-artifact startup

flow, short-context CUDA correctness, the staged KDA lifecycle regression

through 65,536 tokens (snapshot, bit-stable replay, fork, cancellation,

resynchronization), the 2K full-model performance measurement above, and a

measured

131,072-token prefill/decode whose begin/middle/end needles and latest

version record were all retrieved correctly under greedy decoding.

Limitations

• Hardware-specialized deployment. The measured runtime was designed and

tuned for one NVIDIA GB10 with 121.6 GiB coherent unified memory, native

sm_121a kernels, CUDA 13.3, and driver 595.71.05. Memory fit and throughput

do not transfer directly to discrete GPUs, multi-GPU hosts, or other CUDA

architectures.

• Serving-framework dependency. The published files are GGUF weights, but

executing the Solar GQA/KDA family and its compressed runtime KV layouts

depends on the pinned Baekpica/ds4 development branch. Stock llama.cpp,

vLLM, and other GGUF runtimes should not be assumed to expose the same model

family, cache formats, numerical path, or performance.

• Unified-memory assumption. The resident-copy and capacity behavior rely

on GB10's shared CPU/GPU physical memory and its CUDA address-translation

behavior. Conventional host-RAM plus VRAM accounting describes a different

deployment.

• Verified DGX Spark performance covers the runs listed above (2K four-format,

2K-64K lifecycle, 131K prefill/decode with needle retrieval). Full-depth 512K

and 1,048,576-token retrieval semantics are not yet validated, so the tables

are not a full-depth 1M performance claim.

Acknowledgements

• Upstage — Solar Open2 itself, the official

Transformers implementation, and the requirement that derivatives retain

the Solar name and “Built with Solar” attribution.

• antirez/ds4 — the original engine,

loader, session machinery, server, and CUDA architecture this model-family

work extends.

• Entrpi/ds4-on-spark — the DGX

Spark sm_121 foundation and unified-memory CUDA work adopted downstream.

• ggml-org/llama.cpp — GGUF,

llama-quantize, the importance-matrix tooling, and the quantization formats

used by this artifact.

• Upstage's Solar vLLM fork — a useful

reference implementation for separating recurrent decode from chunked

tensor-core KDA prefill.

Errors in the conversion recipe, ds4 model family, and measurements here are

ours, not theirs.

License and attribution

The source model and this derivative artifact are distributed under the

included Upstage Solar License. The derivative retains the Solar name and

this card prominently states “Built with Solar.” Review LICENSE before

use or redistribution.