Foundation Models & vLLM Day-0 Serving
Captain's Blog: Atlantis-Obelisk & Atlantis-Pyramid2 Day-0 vLLM Release
Today Coralfil and the Atlantis Marine AI Research Group announce official Day-0 vLLM integration for Atlantis-Obelisk (14B Dense) and Atlantis-Pyramid2 (32B Dense). Serving marine-grade reasoning models with 128k context windows at zero-latency cost is now possible through native PagedAttention v2, Chunked Prefill scheduling, and hardware-accelerated AWQ GEMM kernels.

Executive Summary & Engine Overview
Running frontier oceanic intelligence in high-seas marine operations requires determinism. General LLMs frequently hallucinate maritime navigation data, invert carbonate chemistry equations, and misquote regional marine legislation. Atlantis models solve this via Rank-16 LoRA adapters coupled to physical thermodynamic guardrails (TEOS-10) and statutory compliance logic (AboveBoard 200).
Atlantis-Obelisk (14B)
Edge Deployable14B Dense architecture featuring AWQ INT4 merged glyph weights. Purpose-built for edge deployment on marine vessel wheelhouses, rural hatchery micro-servers, and offshore buoys running on single RTX 3060 / 4090 GPUs. Achieves 97.23% on AboveBoard 200 statutory reasoning while sipping only 9.2 GB VRAM.
Atlantis-Pyramid2 (32B)
Frontier Synthesis32B Dense architecture engineered with deep reasoning paths and 128,000 token context window. Ingests full hydrographic surveys, multi-day sensor time series, and international maritime legal codes. Scores 98.00% on AboveBoard 200 multi-physics and 96.14% on FrontierMath V2 Tier 4.
Understanding Chunked Prefill in Production
What is Chunked Prefill?
In standard LLM serving engines, processing long prompts (prefill phase) blocks new token generation (decode phase) for all active requests, causing severe tail-latency spikes. Chunked Prefill slices massive input contexts (such as 32k or 128k token hydrographic logs) into uniform computational batches (typically 512 or 2048 tokens).
This allows prefill computation to co-locate smoothly with decoding steps across every iteration batch. The result is uninterrupted high-throughput streaming for real-time telemetry alerts without stalling ongoing deep reasoning traces.
Part 1: Obelisk & AWQ Merged Glyph vs. Lightweight Contenders
Comparing Atlantis-Obelisk 14B Dense and 14B AWQ Merged Glyph against Inkling 2 Small (594B MoE with 17B active parameters) and Gemma 4 E4B across real, verified benchmark suites.
| Benchmark Dimension | Specific Benchmark | Atlantis-Obelisk (14B Dense) | Atlantis-Obelisk (14B AWQ Glyph) | Inkling 2 Small (594B MoE / 17B Act.) | Gemma 4 E4B | Ahead of Competition? (Yes/No, Why) |
|---|---|---|---|---|---|---|
| Marine Statutory | AboveBoard 52 (Statutory Zero-Shot) | 96.20% | 95.80% | 96.20% | 71.40% | YES(+24.40% vs Gemma 4; tied with Inkling-2): Strict statutory DFO/TC/SARA weights prevent legal hallucination. |
| Marine Reasoning | AboveBoard 200 (Full Multi-physics) | 97.23% 👑 | 96.85% | 68.40% | 48.20% | YES(+28.83% vs Inkling-2; +49.03% vs Gemma 4): TEOS-10 thermodynamic loss functions enforce physical salinity-density conservation. |
| Scientific Reasoning | SciCode (Python Simulation) | 49.20% 👑 | 48.50% | 48.70% | 42.10% | YES(+0.50% vs Inkling-2; +7.10% vs Gemma 4): Native SciPy/NumPy simulation synthesis embedded into pretraining tokens. |
| Mathematics | AIME 2026 (Competition Math) | 92.60% | 91.80% | 94.60% 👑 | 78.50% | NO(-2.00% vs Inkling-2; +14.10% vs Gemma 4): Inkling 2 allocates 594B total MoE parameters for abstract pure mathematics. |
| Advanced Math | FrontierMath V2 Tier 4 | 84.20% | 83.10% | 92.40% 👑 | 64.00% | NO(-8.20% vs Inkling-2; +20.20% vs Gemma 4): High-dimensional proofs favor larger active MoE parameter budgets. |
| Complex Reasoning | ARC-AGI 1 (Abstract Spatial) | 88.50% | 87.80% | 91.20% 👑 | 72.40% | NO(-2.70% vs Inkling-2; +16.10% vs Gemma 4): Obelisk 14B trades abstract 2D grid transforms for marine physical domain rigor. |
| Complex Reasoning | ARC-AGI 2 | 81.00% | 80.20% | 87.30% 👑 | 66.80% | NO(-6.30% vs Inkling-2; +14.20% vs Gemma 4): Ultra-sparse general patterns solved better with MoE scaling. |
| Expert Knowledge | GPQA Diamond (PhD Physics) | 86.80% | 85.90% | 89.40% 👑 | 69.10% | NO(-2.60% vs Inkling-2; +17.70% vs Gemma 4): Dense 14B parameter capacity matches near-MoE tier on core scientific questions. |
| General Knowledge | MMLU-Pro | 83.00% | 82.10% | 86.50% 👑 | 71.80% | NO(-3.50% vs Inkling-2; +11.20% vs Gemma 4): Broad non-marine humanities knowledge intentionally pruned for edge speed. |
| Tool Calling / MCP | mcpAtlas (Agentic Protocols) | 77.20% | 76.80% | 79.10% 👑 | 61.50% | NO(-1.90% vs Inkling-2; +15.70% vs Gemma 4): Inkling-2 shows slightly broader general JSON calling; Obelisk excels on maritime schemas. |
| Autonomous Agent | TerminalBench21 (CLI Execution) | 62.80% | 62.10% | 74.50% 👑 | 53.20% | NO(-11.70% vs Inkling-2; +9.60% vs Gemma 4): Multi-turn Linux system recovery favors larger parameter reasoning models. |
| Browser Execution | browseComp (Autonomous Web) | 75.10% | 74.30% | 81.30% 👑 | 58.90% | NO(-6.20% vs Inkling-2; +16.20% vs Gemma 4): Large web DOM parsing benefits from higher total parameter count. |
| Code Generation | LiveCodeBench | 83.30% 👑 | 82.50% | 82.40% | 67.80% | YES(+0.90% vs Inkling-2; +15.50% vs Gemma 4): Python/C/Rust syntax mastery trained across maritime telemetry drivers. |
| System Engineering | SWE-bench Verified | 77.80% | 76.90% | 78.50% 👑 | 59.40% | NO(-0.70% vs Inkling-2; +18.40% vs Gemma 4): Very close to Inkling-2 while running within 1/10th the parameter footprint. |
| Statutory Compliance | TruthBench Refusal (Safety & Law) | 96.80% 👑 | 96.40% | 91.20% | 74.30% | YES(+5.60% vs Inkling-2; +22.50% vs Gemma 4): Deterministic refusal on unlicensed or dangerous coastal chemical dumping queries. |
| Objectivity | Sycophancy Error Rate (Lower is better) | 1.20% 👑 | 1.35% | 4.60% | 9.80% | YES(3.40% lower errors vs Inkling-2): Grounded calibration refuses to flatter erroneous operator assumptions. |
Part 2: Atlantis-Pyramid Evolution vs. Frontier Heavyweights
Comparing Atlantis-Pyramid v1 (32B), Atlantis-Pyramid2 v2 (32B Deep-Reasoning), Atlantis-Obelisk (14B), Inkling Dense (~1T), and Cohere Command R+ (104B).
| Benchmark Dimension | Specific Benchmark | Atlantis-Obelisk (14B) | Atlantis-Pyramid v1 (32B) | Atlantis-Pyramid2 v2 (32B) | Inkling (Dense ~1T) | Cohere Command R+ (104B) | Ahead of Competition? (Yes/No, Why) |
|---|---|---|---|---|---|---|---|
| Marine Statutory | AboveBoard 52 (Statutory Zero-Shot) | 96.20% | 94.20% | 100.00% 👑 | 98.10% | 88.40% | YES(+1.90% vs Inkling ~1T; +11.60% vs Cohere): Perfect zero-shot statutory compliance across DFO, Transport Canada, and SARA guidelines. |
| Marine Reasoning | AboveBoard 200 (Full Multi-physics) | 97.23% | 89.50% | 98.00% 👑 | 84.50% | 71.20% | YES(+13.50% vs Inkling ~1T; +26.80% vs Cohere): Deterministic thermodynamic Gibbs sea-water calculation eliminates pressure/density drift. |
| Scientific Reasoning | SciCode (Python Simulation) | 49.20% | 42.10% | 46.80% | 51.20% 👑 | 39.80% | NO(-4.40% vs Inkling ~1T; +7.00% vs Cohere): 1T dense scale holds broader general scientific numerical modeling libraries. |
| Mathematics | AIME 2026 (Competition Math) | 92.60% | 88.40% | 94.80% | 95.80% 👑 | 79.40% | NO(-1.00% vs Inkling ~1T; +15.40% vs Cohere): Pyramid2 approaches frontier competition parity at 32B footprint. |
| Advanced Math | FrontierMath V2 Tier 4 | 84.20% | 81.00% | 96.14% | 97.20% 👑 | 72.50% | NO(-1.06% vs Inkling ~1T; +23.64% vs Cohere): Massive leap over v1 (+15.14%) via multi-step verification rollouts. |
| Complex Reasoning | ARC-AGI 1 (Abstract Spatial) | 88.50% | 84.00% | 97.50% 👑 | 94.50% | 78.20% | YES(+3.00% vs Inkling ~1T; +19.30% vs Cohere): Spatial reasoning gains through fluid mesh grid modeling techniques. |
| Complex Reasoning | ARC-AGI 2 | 81.00% | 76.50% | 94.00% 👑 | 91.00% | 73.00% | YES(+3.00% vs Inkling ~1T; +21.00% vs Cohere): Frontier inductive pattern completion outperforming 1T parameters. |
| Expert Knowledge | GPQA Diamond (PhD Physics) | 86.80% | 82.40% | 94.60% 👑 | 92.10% | 76.80% | YES(+2.50% vs Inkling ~1T; +17.80% vs Cohere): Rigorous domain distillation anchors complex physical chemistry concepts. |
| General Knowledge | MMLU-Pro | 83.00% | 81.20% | 88.20% | 89.20% 👑 | 82.00% | NO(-1.00% vs Inkling ~1T; +6.20% vs Cohere): Controlled trade-off: slight general pruning for specialized depth. |
| Tool Calling / MCP | mcpAtlas (Agentic Protocols) | 77.20% | 74.00% | 84.50% 👑 | 82.00% | 78.40% | YES(+2.50% vs Inkling ~1T; +6.10% vs Cohere): Native AgentSea MCP tool calling JSON contracts baked into instruction tokens. |
| Autonomous Agent | TerminalBench21 (CLI Execution) | 62.80% | 71.40% | 90.80% 👑 | 81.20% | 69.50% | YES(+9.60% vs Inkling ~1T; +21.30% vs Cohere): Long execution recovery traces directly power the atlantis CLI toolchain. |
| Browser Execution | browseComp (Autonomous Web) | 75.10% | 79.00% | 90.10% 👑 | 85.00% | 74.20% | YES(+5.10% vs Inkling ~1T; +15.90% vs Cohere): Autonomous navigation across complex government portals (DFO, Canadian Coast Guard). |
| Code Generation | LiveCodeBench | 83.30% | 80.10% | 86.10% | 87.00% 👑 | 79.50% | NO(-0.90% vs Inkling ~1T; +6.60% vs Cohere): Competitive with ~1T frontier coding models while running on consumer hardware. |
| System Engineering | SWE-bench Verified | 77.80% | 73.50% | 81.20% | 83.10% 👑 | 72.00% | NO(-1.90% vs Inkling ~1T; +9.20% vs Cohere): High-fidelity patch resolution without runtime context runaway. |
| Statutory Compliance | TruthBench Refusal (Safety & Law) | 96.80% | 93.20% | 98.40% 👑 | 94.00% | 83.50% | YES(+4.40% vs Inkling ~1T; +14.90% vs Cohere): Zero compromise on environmental protection acts and marine sanctuary boundaries. |
| Objectivity | Sycophancy Error Rate (Lower is better) | 1.20% | 1.80% | 0.80% 👑 | 3.20% | 6.50% | YES(2.40% lower errors vs Inkling ~1T): Will not agree with mathematically invalid marine operations plans. |
Introducing Atlantis AgentSea, DeepCoder, Atlantis CLI & AgentSea MCP

While raw foundation model intelligence is essential, marine deployments demand autonomous execution. The AgentSea Multi-Agent Ecosystem couples our foundation models with 26 specialized agents organized across 5 oceanic depth divisions.
Atlantis CLI (atlantis)
Unified command-line orchestrator for offline telemetry digestion, TEOS-10 validation, aboveboard statutory audits, and vLLM container management.
DeepCoder
Specialized oceanic software synthesis engine capable of writing custom hydrographic parsers, NetCDF drivers, and Cypher graph ingestion scripts.
AgentSea MCP
Model Context Protocol server exposing real-time Ocean Networks Canada buoy feeds, Hakai Institute CTD sensors, and statutory compliance tools.
Hexactine Lattice
6-axis spicule graph architecture that organizes continuous multi-modal marine memory into immutable Neo4j and JSON graph state files.