⚠️ This is a gated repository. Model weights are available via two paths — see Deployment Options below.
What Is LATCH
LATCH is a proprietary inference layer built on top of Qwen/Qwen2.5-14B-Instruct that eliminates the long-context performance penalty for document-heavy workloads.
Standard LLMs re-process every document from scratch on every query. LATCH removes this cost entirely — documents are prepared once and subsequent queries run at dramatically reduced latency regardless of document length or count.
This is not RAG. This is not prompt compression. It is a fundamentally different approach to long-context inference that operates at the model level.
Architectural details are proprietary.
Performance Results
All benchmarks run on NVIDIA A100 80GB with vLLM serving infrastructure.
Speed
Metric
Baseline (Qwen 2.5 14B)
LATCH
Improvement
Time-To-First-Token (cold)
23.1s
0.11s
210× faster
TTFT Speedup (avg, customer pack)
4.47s
0.11s
42.9×
End-to-End Query Speedup
6.55s
2.02s
5.2×
Cache Reload Time
23.1s
0.0016s
246× faster
Quality — Customer Document Pack
Benchmark Category
Baseline
LATCH
Delta
Cross-Document Comparison
41.5%
49.4%
+7.9pp
Cross-Document Format
40.5%
68.8%
+28.3pp
Cross-Document Retrieval
40.4%
48.1%
+7.7pp
Selective Retrieval
35.2%
47.2%
+12.0pp
Overall Mean token-F1
39.4%
53.4%
+14.0pp
Benchmark Gates
Gate
Result
Single-Document Gate
11/12 ✅
Multi-Document Gate
11/12 ✅
256K Memory Sweep
Passing
Multi-doc pass rate: 91.7% — the highest of any model family in the current LATCH portfolio.
How It Works
LATCH intercepts the standard inference path and replaces the costly per-query document processing step with a persistent representation that is prepared once and reused across all subsequent queries against the same document set.
The result is a response that begins in under 120 milliseconds — before the user has practically finished pressing Enter — regardless of how many documents are in the corpus.
The underlying method is proprietary and patent pending. CoDynamics Lab does not publish architectural details.
Hardware Requirements
Component
Minimum
Recommended
GPU
NVIDIA A100 40GB
NVIDIA A100 80GB
VRAM
~30 GB
80 GB
CPU RAM
64 GB
128 GB
Storage
50 GB
100 GB
Inference Runtime
vLLM
vLLM ≥ 0.4
LATCH reduces peak VRAM consumption by approximately 50% versus standard Qwen 2.5 14B serving, enabling more concurrent instances per node.
Deployment Options
🔒 Option 1: Self-Hosted License — $79
Run LATCH on your own A100 or H100. Your documents never leave your infrastructure.
Request direct access for evaluation, research, or enterprise licensing discussions.
Intended Use
Primary use cases:
M&A and private equity due diligence (multi-document data room analysis)
Legal document review and cross-contract comparison
Compliance and regulatory document monitoring
Financial research and filing analysis
Any high-volume, repeated-query workload against a fixed document corpus
Out of scope:
Real-time web search or retrieval-augmented generation
General-purpose conversational AI without a document corpus
Consumer applications
Limitations & Known Weaknesses
Short-context standard QA: LATCH is optimized for long-context, multi-document workloads. It does not improve performance on standard short-context QA benchmarks.
Document pre-preparation required: Documents must be prepared before querying. This is a one-time cost per document set that is fully amortized across subsequent queries.
Cross-document retrieval is the weakest benchmark slice: Document-selection tasks with heavy distractors are the most challenging workload category.
Request Access
Three ways to get started:
Path
Best for
Action
Self-hosted license
Teams with their own A100/H100 who need full data privacy
