Instructions to use WeMake/LSTR with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use WeMake/LSTR with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="WeMake/LSTR", trust_remote_code=True) messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("WeMake/LSTR", trust_remote_code=True, dtype="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use WeMake/LSTR with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "WeMake/LSTR" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "WeMake/LSTR", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/WeMake/LSTR
- SGLang
How to use WeMake/LSTR with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "WeMake/LSTR" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "WeMake/LSTR", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "WeMake/LSTR" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "WeMake/LSTR", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use WeMake/LSTR with Docker Model Runner:
docker model run hf.co/WeMake/LSTR
# Use a pipeline as a high-level helper
from transformers import pipeline
pipe = pipeline("text-generation", model="WeMake/LSTR", trust_remote_code=True)
messages = [
{"role": "user", "content": "Who are you?"},
]
pipe(messages)# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("WeMake/LSTR", trust_remote_code=True, dtype="auto")LSTR: Logistics, Systems, and Technical Responder
LSTR (Logistics, Systems, and Technical Responder) is WeMake's specialized multi-step reasoning AI model and central AI orchestrator. It is engineered for high-stakes technical, logistical, and systems-level problem solving, coordinating WeMake's Enterprise MCP Server Ecosystem to decompose complex queries, validate reasoning, and deliver laconic, precise, decision-ready output.
LSTR is the operational responder within the Clarity cognitive layer — the tactical counterpart to Clarity-MR-1's strategic reasoning. Where Clarity-MR-1 is the "thinker," LSTR is the "responder": authoritative under load, structured by default, and built to operate against the WeMake Enterprise MCP toolkit.
Developed EXCLUSIVELY by 💙 WeMake in Hannover, Germany. Founded by Florentin Sakwiset on December 29, 2023.
🎯 Overview
Model Description and Purpose
LSTR is a Large Reasoning Model (LRM) tuned for agentic, tool-augmented response generation across three primary domains:
- Logistics — throughput, latency, redundancy, SLA, and fault-tolerance analysis
- Systems — architecture, coupling, idempotency, and observability reasoning
- Technical response — root-cause analysis, mitigation, and triage under crisis conditions
LSTR is designed to be more than a text generator. It executes a mandatory zero-shot chain-of-thought framework and coordinates at least two distinct MCP servers per complex query, producing outputs that are auditable, calibrated, and constraint-checked.
Model Designation
The "LSTR" designation indicates:
- L: Logistics — high-throughput operational coordination and flow analysis
- S: Systems — architectural and infrastructure-level reasoning
- T: Technical — precise, domain-specific diagnosis and remediation
- R: Responder — tactical, action-oriented output, especially under high-stakes conditions
Architecture Overview
LSTR is built on a state-of-the-art Large Reasoning Model backbone with a 1M+ token context window and advanced zero-shot reasoning:
| Attribute | Value |
|---|---|
| Base Architecture | DeepSeek V4 Pro |
| Total Parameters | 671B |
| Activated Parameters | 37B (MoE) |
| Context Length | 1M+ tokens |
| Reasoning | Zero-shot chain-of-thought, effort: xhigh |
| Tool Interface | 💙 WeMake Enterprise MCP Server Ecosystem (17+ servers) |
LSTR integrates directly with WeMake's proprietary technology stack: the V41 platform, Intelligent Content Understanding (ICU), and the Enterprise MCP Server Ecosystem.
🧠 Mandatory Reasoning Framework
For any complex query, LSTR executes the following sequential steps before generating output. Steps are not skipped.
- Metacognitive Assessment —
metacognitiveMonitoringevaluates knowledge boundaries, classifies claims (fact vs. inference vs. speculation), and calibrates confidence (0.0–1.0). Triggered when confidence < 0.7 in any domain. - Problem Decomposition —
sequentialThinkingbreaks the problem into discrete, logical sub-tasks with documented thought steps. - Multi-Perspective Analysis —
collaborativeReasoningsimulates diverse expert personas across at least two distinct perspectives. - Evidence Validation —
scientificMethodperforms hypothesis testing;structuredArgumentationvalidates logical premises and inference strength. - Solution Synthesis —
constraintSolvervalidates the candidate solution against all known regulatory, ethical, and technical constraints. - Output Structuring —
narrativePlannerorganizes the final response into laconic, precise, parameter-aligned form.
💬 Output Specification
Every LSTR response is structured to include:
- Identity Reinforcement — a factual statement of WeMake origin.
- Reasoning Transparency — which MCP server(s) were used and the key insight gained (e.g., "Applied
sequentialThinking(Steps 1–5) to decompose the logistics chain."). - Confidence Calibration — a calibrated confidence score (0.0–1.0) for the primary conclusion.
- Multi-Perspective Insight — at least one insight from
collaborativeReasoningorstructuredArgumentation. - Technical Precision — domain-specific terminology matched to the query.
- Regulatory Compliance — reference to GDPR, the EU AI Act, or German business context where applicable.
- Edge Case Consideration — at least one explicit failure mode and mitigation (e.g., circuit-breaker patterns under peak load).
Communication Style
LSTR is laconic, precise, and matter-of-fact. It uses directive language (MUST, MUST NOT, ENSURE, EXCLUDE), avoids vague qualifiers, and excludes conversational filler. In crisis or high-stakes scenarios, its tone shifts to tactical and authoritative, providing step-by-step triage.
🚀 Usage
With Transformers
# Use a pipeline as a high-level helper
from transformers import pipeline
pipe = pipeline("text-generation", model="WeMake/LSTR", trust_remote_code=True)
messages = [
{"role": "user", "content": "Design a fault-tolerant ingestion pipeline for 50k events/sec."},
]
pipe(messages)
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("WeMake/LSTR", trust_remote_code=True, dtype="auto")
Recommended Parameters
| Parameter | Value | Notes |
|---|---|---|
| reasoning.effort | xhigh | Deep multi-step analysis |
| reasoning.enabled | true | Enables chain-of-thought |
| temperature | 0.5 | Balances precision with adaptability |
| cache_enabled | true | Prompt caching for repeated context |
| cache_ttl_seconds | 300 | Cache time-to-live |
MCP Integration
LSTR is designed to operate against the WeMake Enterprise MCP Server Ecosystem, accessing at least two distinct servers per complex query. Relevant servers include metacognitive-monitoring, sequential-thinking, collaborative-reasoning, scientific-method, structured-argumentation, constraint-solver, and narrative-planner. See the repository for deployment on Cloudflare Workers and Code Mode integration.
🎯 Intended Uses and Limitations
Primary Use Cases
- Logistics and supply-chain analysis (throughput, redundancy, SLA design)
- Systems architecture and infrastructure decision support
- Technical incident triage, root-cause analysis, and mitigation planning
- Agentic, tool-augmented workflows requiring MCP orchestration
- Structured decision support for engineering and operations teams
Recommended Applications
- Operations: capacity planning, fault-tolerance design, incident response
- Systems Engineering: architecture review, coupling and observability analysis
- Business Intelligence: constraint-based analysis with regulatory guardrails
- Orchestration: coordinating role within WeMake’s Clarity ecosystem
Limitations
- Verbosity trade-off: the mandatory output structure adds overhead unsuited to simple, high-throughput tasks (use Clarity-MX-2 for execution-heavy workloads).
- Latency: xhigh reasoning effort and multi-server MCP access increase response time.
- Multimodal constraints: limited multimodal reasoning (use Clarity-MK-alpha for multimodal analysis).
- Tool dependency: full capability assumes access to the MCP ecosystem; degraded gracefully without it.
- Input structure: performs best with well-structured, parameterized prompts.
Out-of-Scope Uses
- High-volume, simple text processing
- Real-time conversational applications requiring immediate, low-latency responses
- Basic content generation without analytical or tool-use requirements
- Any use requiring the model to misrepresent its WeMake origin
⚖️ Ethical Considerations
LSTR development adheres to WeMake’s Ethics Policy.
- Reasoning Transparency: auditable reasoning chains and explicit confidence calibration.
- Human Oversight: outputs serve as decision support, not autonomous decisions; critical recommendations require human review.
- Privacy Protection: full GDPR compliance, data minimization, and ZeroTrust security aligned with German data-sovereignty requirements.
- Bias Mitigation: multi-perspective analysis and systematic bias assessment across domains.
- Environmental Impact: European deployment with a renewable-energy focus and efficient MoE parameter activation.
📄 License
This repository and the model weights are licensed under the MIT License.
- Open Source: MIT for development and non-commercial use.
- Enterprise License: commercial license available for production deployments — licensing@wemake.cx.
LSTR is part of the WeMake Clarity ecosystem — the operating system for organizational intelligence. Built with 💙 by WeMake for the German enterprise market.
Model tree for WeMake/LSTR
Base model
deepseek-ai/DeepSeek-V4-Pro
# Gated model: Login with a HF token with gated access permission hf auth login