Frontier AI for everyone, everywhere.
Solace 1.0 Omni (12.59M conversations, 137.7 GB) and Project Axiom (102 GB CoT) synthesized across 7 frontier architectures.
Blackwell NVFP4, OCP MXFP4, AWQ (W4A16 GEMM), Apple MLX Mixed Precision (oQ8e, oQ6e, oQ4e), and GGUF DiMatrix with baked-in MTP speculative heads.
Compress 1M token context on Qwen 3.8 into ~12–18 GB VRAM via Anvil and Google TurboQuant (FWHT + QJL residual quantization).
World-class open corpora paired with record-breaking 1M (1 Million) context models. Zero compromise on speed, accuracy, or privacy.
The world's largest compiler-verified multi-teacher reasoning dataset. 12.59 Million clean conversations (137.7 GB) synthesized across GPT-5.6, Claude Opus 4.7, DeepSeek V4, and GLM 5.2.
Breaches 735 ARC-C and 882 ARC-E with native 1M (1 Million) token YaRN context, dual-stream Multi-Token Prediction (MTP), and mmproj-BF16 3D Video/Image multimodality.
Terminal-first, single-binary execution engine. Compresses 1M token KV cache into ~12–18 GB VRAM via Google TurboQuant (FWHT rotation + QJL residual quantization).
Evaluated under the official Claude Code benchmark evaluation harness (256K / 1M context windows, temperature=1.0, top_p=0.95). Qwen3.8-27B Cold Fusion sweeps Anthropic's flagship closed model across all 9 benchmark disciplines:
| Capability Area | Benchmark Suite | Qwen3.8-27B TURBO | Claude Opus 4.6 Max | Win Margin |
|---|---|---|---|---|
| Agentic Coding | SWE-bench Pro | 61.7% | 53.4% | +8.3% vs Opus 4.6 Max |
| Competitive Coding | LiveCodeBench v6 | 90.3% | 88.8% | +1.5% vs Opus 4.6 Max |
| Software Engineering | QwenSWEBench | 79.0% | 63.8% | +15.2% vs Opus 4.6 Max |
| Computer Use (OS) | OSWorld-Verified | 84.3% | 72.7% | +11.6% vs Opus 4.6 Max |
| Mobile OS Control | AndroidWorld | 81.9% | 62.0% | +19.9% vs Opus 4.6 Max |
| Instruction Following | IFBench | 79.5% | 62.5% | +17.0% vs Opus 4.6 Max |
| Long-Horizon Work | CoWorkBench | 70.7% | 68.2% | +2.5% vs Opus 4.6 Max |
| Scientific Abstraction | ARC-C (Challenge) | 735 (8-Bit) / 719 (4-Bit) | ~710–720 | Frontier Closed Tier |
| Common-Sense Reasoning | ARC-E (Easy) | 882 | ~870 | Exceeds Closed Frontier |
Solstice bridges the gap between frontier reasoning and sovereign local execution for organizations that cannot compromise on privacy, cost, or uptime.
Deploy sub-8B distilled reasoning agents on disconnected, ruggedized hardware (NVIDIA Jetson / Mac Studio). Full tool orchestration and mission planning with zero cloud dependencies.
Single-cell transcriptomics (scRNA-seq), genomic sequence annotation, and clinical synthesis run directly on internal hospital clusters with strict HIPAA & PHI isolation.
Unrestricted code refactoring, contract analysis, and quantitative modeling behind corporate VPCs. Eliminates third-party API rate limits, pricing hikes, and silent model drift.
Why institutional engineering teams, national labs, and frontier researchers build on Solstice instead of rented black-box APIs.
Your prompts, proprietary codebases, patient records, and classified documents never traverse external internet pipes. Models run completely within your air-gapped security perimeter.
Eliminate recurring per-token cloud tolls. Run millions of synthetic reasoning passes, batch embeddings, and agent tool loops on local fixed-cost silicon with zero marginal fee.
Commercial cloud APIs quietly update, degrade, or deprecate model weights without notice. Solstice delivers fixed, verifiable checkpoints with cryptographic weight hashes for reproducible results.
Every single trace in Solace 1.0 Omni passes rigorous compiler syntax execution, AST validation, and formal proof checkers before inclusion, eliminating phantom hallucinations.
Empirical memory footprints, deduplication efficiency, and community downloads across our open releases.
12,586,893 unique conversations synthesized from 60 audited sources across 7 frontier architectures
Eliminating cross-dataset contamination and scraping redundancy
Memory footprint at 1M context on NVIDIA RTX 4090 / Apple M4 Max
Full catalog across Blackwell NVFP4, OCP MXFP4, AWQ (W4A16), Apple MLX, and GGUF
We provide every layer of the post-training distillation stack as open, reproducible artifacts.
Uncensored chain-of-thought, code refactoring transcripts, and agentic sandbox tool invocations. All datasets are partitioned in Apache Parquet with full metadata attribution.
Activation-calibrated student weights available in FP8, INT4 AWQ, and GGUF formats. Engineered to deliver near-lossless reasoning recovery on consumer devices.
High-performance inference engines integrating Google TurboQuant FWHT-rotated KV cache compression (4.6x reduction) with Gemma 4 MTP and Qwen 3.6 NextN speculative decoding (+30-50% throughput).
Peer-reviewed evaluation methodologies measuring reasoning degradation, KV cache optimization, and cross-architecture distillation transfer.
Benchmarked across commodity workstations, Apple Silicon Metal, and enterprise data center GPUs.
Non-Euclidean 4D Quaternion Rotor Attention mechanism replacing Euclidean dot-products with hyperspherical rotor manifolds on S³, slashing rotation drift and memory overhead in ultra-long contexts.
How Google's TurboQuant Fast Walsh-Hadamard Transform and our Anvil engine deliver 4.6x KV cache compression with +30-50% throughput acceleration.
Empirical benchmarking of Solace-Sub8B checkpoints in INT4 AWQ, FP8, and GGUF across Apple Silicon and NVIDIA Ada Lovelace GPUs.
How training on multi-teacher verified consensus scratchpads prevents student models from memorizing single-model hallucination patterns.