Instructions to use Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF
# Run inference directly in the terminal:
llama cli -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF
# Run inference directly in the terminal:
llama cli -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF
# Run inference directly in the terminal:
./llama-cli -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

Use Docker

docker model run hf.co/Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

LM Studio
Jan

vLLM

How to use Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

Ollama
How to use Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF with Ollama:
```
ollama run hf.co/Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF
```
Unsloth Desktop

How to use Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF with Docker Model Runner:
```
docker model run hf.co/Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF
```

Lemonade

How to use Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

Run and chat with the model

lemonade run user.Solar-Open2-250B-Mixed-Quant-GGUF-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Baekpica commited on 28 days ago

Commit

bf141d7

verified ·

1 Parent(s): fcc06ed

Add mixed quantization report

Browse files

Files changed (1) hide show

MIXED-QUANT.md +105 -0

MIXED-QUANT.md ADDED Viewed

	@@ -0,0 +1,105 @@

+# Solar Open 2 250B MXQ-v1 artifact report
+Built with Solar.
+## Result
+`MXQ-v1` is the single locked mixed-quantization candidate for
+`upstage/Solar-Open2-250B@9190fbe63a2ad8e17fc766ccceb36de7c66f004b`.
+It preserves all 48 layers, all 320 routed experts per layer, the shared
+expert, and the original hybrid GQA/KDA schedule.
+The final artifact is 11 GGUF shards totaling `95,533,532,160` bytes
+(`88.972535133362 GiB`). This is `3.027465 GiB` below the `92 GiB` hard
+limit. Static verification passed for all `1,083` GGUF tensors. Per-shard
+SHA-256 values are recorded in `manifests/sha256-v1.txt`.
+## Locked target precision schedule
+| Target module family | Precision |
+|---|---|
+| Token embedding and LM head | `Q8_0` |
+| GQA q/k/v/o/g projections | `Q8_0` |
+| KDA matrix projections | `Q8_0` |
+| Shared expert gate/up/down | `Q8_0` |
+| Routed expert gate/up, middle layers 4–43 | `IQ2_XXS` + imatrix |
+| Routed expert down, middle layers 4–43 | `Q3_K` + imatrix |
+| Routed expert gate/up/down, edge layers 0–3 and 44–47 | `Q4_K` + imatrix |
+| Router weights/correction bias | `F32` |
+| Norms and small KDA control/norm/depthwise-convolution tensors | `F32` |
+No lower-precision fallback candidate was generated. The complete override
+list and machine-readable recipe are in `manifests/quant-overrides-v1.txt`
+and `manifests/quant-recipe-v1.yaml`.
+The verified tensor-type distribution is:
+| GGUF type | Tensor count |
+|---|---:|
+| `F32` | 409 |
+| `Q8_0` | 530 |
+| `IQ2_XXS` | 80 |
+| `Q3_K` | 40 |
+| `Q4_K` | 24 |
+## Calibration and imatrix
+The calibration mix follows the published target token shares from
+`Baekpica/Solar-Open2-120B-A15B-REAM-148E-Healing-Mix@1931f3a40cc3463217f9c7d25906f80ded029264`:
+| Bucket | Token share |
+|---|---:|
+| Instruction-following chat | 22% |
+| Cascade stage 1 reasoning | 16% |
+| Cascade stage 2 reasoning | 16% |
+| Korean | 16% |
+| Other multilingual | 12% |
+| Finance | 6% |
+| SWE agentic | 6% |
+| Algorithmic code | 6% |
+The rendered corpus contains 1,641 documents, 4,056,882 official-tokenizer
+tokens, and 15,401,547 bytes. The final imatrix processed 590 chunks of 512
+tokens (`302,080` tokens) and is `590,796,096` bytes with SHA-256
+`1ee8f4d69e9e7478f9ff23159e0a08f2306f176e9d52cf85bac38151963a56e7`.
+Coverage is accepted with one explicit exception: layer 0 has 29 of 320
+routed experts with zero activation in this pinned corpus. Layers 1–47 have
+strict 320/320 coverage. The verifier rejects any larger layer-0 exception or
+any zero-coverage expert in layers 1–47; the exact expert IDs are retained in
+`manifests/verify-imatrix.json`.
+## H100 verification
+- Quantization completed in `3,021.266` seconds (`50m21s`).
+- Structural verification found zero errors across all 11 shards.
+- llama.cpp loaded the artifact over four H100 NVLs and generated all quality
+  fixtures; the observed decode rate was about 71 tokens/s and is reported as
+  H100 development evidence only.
+- All 32 short deterministic outputs were non-empty with valid token IDs;
+  versus the full-model `Q8_0` reference, 26/32 first tokens and 11/32 complete
+  32-token sequences matched.
+- All eight longer no-think fatal/content/format sanity checks passed,
+  including Korean and Japanese answers, arithmetic, typed code, exact JSON,
+  and list-format constraints.
+- ds4 mapped the complete artifact on one H100, validated all mixed weights,
+  and passed snapshot/restore, recurrent-state, replay, and cold-rebuild
+  regressions. See `reports/CUDA-KDA.md` for numerical details.
+These checks establish artifact integrity and H100 execution sanity. They are
+not a benchmark-grade semantic evaluation and do not establish DGX Spark
+compatibility or one-million-token serving.
+## Reproduction gates
+```bash
+scripts/verify_imatrix.sh
+scripts/verify_artifact.sh v1
+scripts/run_v1_quality.sh
+CUDA_VISIBLE_DEVICES=0 \
+  DS4_SOLAR_MODEL=/path/to/Solar-Open2-250B-MXQ-v1-00001-of-00011.gguf \
+  JOBS=32 scripts/run_h100_regression.sh
+```
+Pinned inputs, exact projections, quality summaries, and raw H100 logs are in
+`manifests/`, `fixtures/reference/`, and `benchmarks/h100/`.