Quantized Flagship Checkpoints Catalog

Available Flagship Checkpoints

All Solstice-distilled model checkpoints and quantizations are published under the official Solstice-AI organization on Hugging Face:

1. Qwopus3.8-27B-Flash Suite (1M Native Context)

Model Identifier	Format / Precision	Context	Key Features	Target Hardware	Compatible Engine
Qwopus3.8-27B-Flash-AWQ-1M	AWQ INT4 (W4A16 GEMM)	1M Tokens	AutoRound W4A16, 1M YaRN scaling	RTX 3090 / 4090, A100, L40S	vLLM / SGLang
Qwopus3.8-27B-Flash-GGUF-1M	GGUF Dynamic v3.0	1M Tokens	UD-IQ4_XS to UD-Q8_XL + mmproj-BF16	CPU, CUDA, Apple Metal	llama.cpp
Qwopus3.8-27B-Flash-mlx-oQ4e-1M	MLX 4-Bit Mixed	1M Tokens	Apple Metal MTP, 2.2x speed	Apple Silicon (16GB–24GB RAM)	Apple MLX
Qwopus3.8-27B-Flash-mlx-oQ8e-1M	MLX 8-Bit Mixed	1M Tokens	Apple Metal MTP, reference grade	Apple Silicon (36GB+ RAM)	Apple MLX
Qwopus3.8-27B-Flash-NVFP4-1M	Blackwell NVFP4	1M Tokens	Native NVFP4 Tensor Cores + MTP	RTX 4090 / 5090, L40S	TensorRT-LLM / vLLM
Qwopus3.8-27B-Flash-MXFP4-1M	OCP MXFP4	1M Tokens	Microscaling FP4 + MTP	Universal 24GB+ GPU	vLLM / SGLang

2. DeepSeek-V4-Flash-Vision Suite (Multimodal & Uncensored)

Model Identifier	Format / Precision	Context	Key Features	Target Hardware	Compatible Engine
DeepSeek-V4-Flash-Vision-UNCENSORED-AWQ-DSpark	AWQ INT4 (W4A16 GEMM)	Native	High-speed spatial reasoning + DSpark	RTX 3090 / 4090, A100	vLLM / SGLang
DeepSeek-V4-Flash-Vision-UNCENSORED-GGUF-DSpark	GGUF Matrix + mmproj	Native	Multimodal vision with DSpark drafter	CPU, CUDA, Metal	llama.cpp
DeepSeek-V4-Flash-Vision-UNCENSORED-mlx-oQ4e-DSpark	MLX 4-Bit Mixed	Native	High-efficiency Apple Metal vision	Apple Silicon (16GB–24GB RAM)	Apple MLX
DeepSeek-V4-Flash-Vision-UNCENSORED-mlx-oQ8e-DSpark	MLX 8-Bit Mixed	Native	High-precision Apple Metal vision	Apple Silicon (36GB+ RAM)	Apple MLX
DeepSeek-V4-Flash-Vision-UNCENSORED-NVFP4-DSpark	Blackwell NVFP4	Native	Native Tensor Core FP4 acceleration	RTX 4090 / 5090, L40S	TensorRT-LLM / vLLM
DeepSeek-V4-Flash-Vision-UNCENSORED-MXFP4-DSpark	OCP MXFP4	Native	Microscaling FP4 with DSpark drafter	Universal 24GB+ GPU	vLLM / SGLang

3. GLM-5.3-Flash Suite (High-Throughput Uncensored)

Model Identifier	Format / Precision	Context	Key Features	Target Hardware	Compatible Engine
GLM-5.3-Flash-UNCENSORED-AWQ	AWQ INT4 (W4A16 GEMM)	Native	Low-latency agent rollouts	RTX 3090 / 4090, A100	vLLM / SGLang
GLM-5.3-Flash-UNCENSORED-GGUF	GGUF Quantized Matrix	Native	Comprehensive desktop deployment	CPU, CUDA, Metal	llama.cpp
GLM-5.3-Flash-UNCENSORED-NVFP4-DSpark	Blackwell NVFP4	Native	High-throughput batch reasoning	RTX 4090 / 5090, L40S	TensorRT-LLM / vLLM
GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DSpark	MLX 4-Bit Mixed	Native	Compact mobile workstation inference	Apple Silicon (16GB RAM)	Apple MLX
GLM-5.3-Flash-UNCENSORED-mlx-oQ8e-DSpark	MLX 8-Bit Mixed	Native	Lossless reference agent trace grade	Apple Silicon (32GB+ RAM)	Apple MLX

4. Qwen3.8-27B-TURBO NEO-CODER Suite (Agentic SWE)

Model Identifier	Format / Precision	Context	Key Features	Target Hardware	Compatible Engine
NEO-CODER-AWQ-1M	AWQ INT4 (W4A16 GEMM)	1M Tokens	Full repository SWE-bench verification	RTX 3090 / 4090, A100	vLLM / SGLang
NEO-CODER-NVFP4-1M	Blackwell NVFP4	1M Tokens	Microscaling Tensor Cores + MTP	RTX 4090 / 5090, L40S	TensorRT-LLM / vLLM
NEO-CODER-MXFP4-1M	OCP MXFP4	1M Tokens	Universal enterprise cluster serving	Universal 24GB+ GPU	vLLM / SGLang
NEO-CODER-GGUF	GGUF Quantized Matrix	262K Tokens	Integrated speculative drafting	CPU, CUDA, Metal	llama.cpp
NEO-CODER-mlx-oQ4e-1M	MLX 4-Bit Mixed	1M Tokens	Local agentic coding on macOS	Apple Silicon (16GB–24GB RAM)	Apple MLX
NEO-CODER-mlx-oQ8e-1M	MLX 8-Bit Mixed	1M Tokens	Reference competitive coding grade	Apple Silicon (36GB+ RAM)	Apple MLX

---

Quantization Formats & Specifications

AWQ (Activation-aware Weight Quantization)

• Format: INT4 weights, FP16 activations (W4A16 GEMM).
• Footprint: ~18–20 GB total weights for 27B parameter models.
• Design Target: Production GPU serving engines including vLLM and SGLang. Preserves salient weight channels by analyzing activation magnitude distributions, preventing outlier clipping degradation in reasoning tasks.

GGUF Dynamic Matrix (Unsloth Dynamic v3.0)

• Format: Mixed per-tensor quantization ranging from UD-IQ4_XS (compact desktop footprint) to UD-Q8_XL (lossless reference quality).
• Vision Projection: Packaged with companion mmproj-BF16.gguf for visual spatial-temporal comprehension on vision-capable architectures.
• Design Target: Local deployment across CPU, CUDA, and Apple Silicon unified memory via standard llama.cpp bindings.

Apple MLX Mixed Precision

• Format: 4-bit (mlx-oQ4e), 6-bit (mlx-oQ6e), and 8-bit (mlx-oQ8e) mixed-precision configurations.
• Footprint: ~17.0 GB for oQ4e (operational on 16GB–24GB unified memory MacBooks) and ~30.0 GB for oQ8e.
• Design Target: Native Apple Silicon Metal acceleration with integrated Multi-Token Prediction (MTP) drafting support.

NVIDIA NVFP4 & OCP MXFP4

• Format: 4-bit floating point microscaling format designed for Blackwell and Ada Lovelace Tensor Cores.
• Footprint: ~18.8 GB for NVFP4, ~18.0 GB for OCP MXFP4.
• Design Target: Next-generation datacenter and workstation silicon optimizing FP4 arithmetic throughput while maintaining activation fidelity.