You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Elster-1 is a concept release. Describe your role (technician, scheduler, researcher, clinician, farmer, knowledge worker) and the node of your value chain you want to augment.
Log in or Sign Up to review the conditions and access this model content.
Elster-1: Augmented Reasoning Along the German Value Chain
Status: Concept (v0.1, 29.10.2026). This card specifies a model that has not been trained yet. Architecture figures are inherited from Kolibri-1 and are real. Every Elster-1 result is TBA until measured. We do not publish estimates as results.
Elster-1 is Florentin One's German-first reasoning model for augmentation, not automation. It is a finetune of Aleph-Alpha/Kolibri-1 that moves the cognitive discipline of the LSTR harness (the six-step reasoning chain and seven "cognitive organs") from external MCP tool calls into the model weights. The goal is that a technician, scheduler, researcher, clinician, farmer, or knowledge worker gets structured, calibrated, auditable thinking support on their own hardware, with the human remaining the final authority (der Mensch bleibt die letzte Instanz).
Executive Summary
Elster-1 makes the Klarheit Manifest concrete: AI that leaves people better at thinking, not more dependent on the tool. It inherits Kolibri-1's native German, small active footprint (3.46B active / 78.1B total), and trained abstention. It adds seven reasoning capabilities that LSTR currently has to run through a remote MCP server.
Disclosure principle: we state where Elster-1 is weak before we state where it is strong.
Not for
- Not for autonomous decisions; every output is decision support for a named human role
- Not for medical diagnosis or therapy decisions (that would make it a medical device under MDR, high-risk under EU AI Act Annex I)
- Not for evaluating, ranking, or allocating work to people based on their behaviour or traits (EU AI Act Annex III(4)); scheduling covers machines, material, and orders, not personnel performance
- Not for predictive maintenance, HR, marketing, or customer support (out of scope for the BMFTR value-chain programme this model targets)
- Not for image, sensor, or audio input; Elster-1 is text-only like its base
- Not for languages other than German and English
- Not for real-time control loops (PLC/OT); Elster-1 advises operators and does not actuate anything
Details: see Limitations and Out-of-Scope Uses.
🎯 Overview
Model Designation
Elster (German for magpie) follows Kolibri's bird lineage. The Eurasian magpie is the first non-mammal shown to recognise itself in a mirror (Prior, Schwarz & Güntürkün, PLoS Biology, 2008). That is the design premise: a model that sees its own reasoning before it hands it to a person.
- E: Elster, the reflective German companion to Kolibri
- 1: first generation; capabilities distilled from LSTR, weights from Kolibri-1
Lineage
| Layer | Source | What Elster-1 takes from it |
|---|---|---|
| Weights and language | Kolibri-1 (Aleph Alpha) | Bilingual DE/EN MoE, UniBPE German tokenizer, Merlin-Arthur abstention, reasoning effort levels, on-prem footprint |
| Reasoning discipline | LSTR harness + LSTR-1/reasoning | Six-step chain, seven cognitive organs, complexity gate, "confidence is never invented" rule, deterministic graders |
| Purpose and ethics | Klarheit Manifest | Augmentation over automation, cognitive-debt awareness, lean method along value chains, the human as final authority |
| Card conventions | Clarity-MR-1, LSTR | "Not for" box first, disclosure principle, TBA instead of invented numbers |
Why a finetune instead of the harness alone
LSTR runs its reasoning chain by calling mcp.api.florentin-one.de. That has three costs: a network dependency, a GDPR Art. 28 processing relationship for every reasoning call, and latency for each of the 6+ tool calls. Elster-1 internalises the chain so it runs locally by default. The MCP tools remain available as an optional audit mode that produces an external, machine-checkable reasoning record.
🧠 The Seven Capabilities
Each capability maps one-to-one to a Florentin One cognitive organ. In Elster-1 it is a trained behaviour, not a tool call.
| # | Capability | Cognitive organ (MCP origin) | What the user sees | Trained behaviour |
|---|---|---|---|---|
| 1 | Self-awareness (Selbstwahrnehmung) | metacognitiveMonitoring |
Every key claim is tagged fact / inference / speculation, with a confidence from 0.0 to 1.0 | Calibrated confidence; abstains when context does not support an answer (Kolibri Merlin-Arthur base) |
| 2 | Structured thinking (Strukturiertes Denken) | sequentialthinking |
The problem broken into ≥3 ordered steps with explicit dependencies | Decomposition with revision and backtracking |
| 3 | Multi-perspective thinking (Mehrperspektivisches Denken) | collaborativeReasoning |
≥2 genuinely different viewpoints (e.g. operator vs. quality vs. regulator) | Persona-divergence training; penalises "perspectives" that only restate each other |
| 4 | Hypothesis testing (Hypothesenprüfung) | scientificMethod |
A testable hypothesis, what would falsify it, and the next check to run | Falsification-first reasoning; proposes a check instead of asserting a cause |
| 5 | Dialectical thinking (Dialektisches Denken) | structuredArgumentation |
Thesis, antithesis, synthesis for contested choices | Steelmanning the opposing position before resolving |
| 6 | Boundary regulation (Grenzregulation) | constraintSolver |
Hard limits checked explicitly: numeric, regulatory, safety, scope | Numeric constraints verified; escalates to a human when a boundary is unclear or crossed |
| 7 | Meaning-making (Sinnkonstruktion) | narrativePlanner (re-scoped) |
One short closing section: what this means for your next decision, in your role's language | Ties the result to the user's goal and the next handoff in the chain |
Design note on meaning-making. The LSTR harness deliberately excludes
narrativePlannerfrom analytical output because it produces three-act story outlines. Elster-1 keeps the capability but scopes it narrowly: meaning-making explains relevance and the next action. It MUST NOT impose a narrative arc, suspense, or persuasion on analytical content.
Complexity gate (inherited from LSTR)
The full seven-capability chain runs only when two or more of these hold: multi-step inference, competing constraints, non-trivial uncertainty, material consequence, explicit user demand. Simple questions get direct answers. This keeps cost proportional to the problem, so the reasoning scaffold is not imposed on every request.
Augmentation modes (against cognitive debt)
Kosmyna et al. (2025) measured weaker neural, linguistic, and behavioural performance in LLM-assisted writers over four months. Elster-1 therefore lets the deployment decide how much thinking the model does for the user:
| Mode | Behaviour | Typical use |
|---|---|---|
coach |
Asks guiding questions, shows the structure, and withholds the answer until the user proposes one | Training, onboarding, apprentices (Azubis), skill retention |
copilot (default) |
Full reasoning shown; user confirms each consequential step | Daily work |
responder |
Concise answer with confidence and the one most important caveat | Time-critical triage by experienced staff |
🔗 Roles Along the Value Chain
Elster-1 targets nodes of a vertical chain, not a horizontal chatbot. The example below shows one chain (agri-food to processing) to illustrate the handoffs. Other chains (manufacturing, chemistry, medtech) follow the same pattern.
| Role | Value-chain node | Cognitive bottleneck | Lead capabilities | Human stays decider on |
|---|---|---|---|---|
| Farmer (Landwirt:in) | Primary production | Nutrient planning across soil data, weather, and the Düngeverordnung | 6 Boundary regulation, 4 Hypothesis testing | Application rates and timing |
| Researcher (Forscher:in) | Lab / R&D | Designing experiments that can actually fail; literature synthesis | 4 Hypothesis testing, 5 Dialectics, 1 Self-awareness | Study design, conclusions |
| Scheduler (Disponent:in) | Production and logistics planning | Competing constraints: capacity, deadlines, changeovers | 6 Boundary regulation, 2 Structured thinking, 3 Perspectives | Final plan and trade-offs |
| Technician (Techniker:in) | Commissioning, troubleshooting, repair | Fault isolation from manuals, logs, and symptoms | 2 Structured thinking, 4 Hypothesis testing, 6 Safety boundaries | Every physical intervention |
| Clinician (Ärzt:in) | Medical documentation and guideline work | Guideline lookup, letter drafting, evidence summaries | 1 Self-awareness, 3 Perspectives, 7 Meaning-making | Diagnosis and therapy (out of scope for the model) |
| Knowledge worker (Wissensarbeiter:in) | Synthesis and coordination across the chain | Tool fragmentation, context switching | 2 Structured thinking, 7 Meaning-making | What gets decided and sent |
Handoff behaviour. When output is meant for the next node (e.g. a researcher's finding that a scheduler must act on), Elster-1 can emit a structured handoff block: claim, confidence, assumptions, open checks, and the decision the recipient has to make.
📐 Model Overview
| Model | Elster-1 |
|---|---|
| Model Provider | Florentin One, Hannover, Germany |
| Base Model | Aleph-Alpha/Kolibri-1 (finetune; no architecture changes) |
| Architecture | Mixture-of-Experts, 50 layers, 384 routed + 1 shared expert, 6 routed active (inherited) |
| Total parameters | 78.1B (inherited) |
| Active parameters / token | 3.46B (inherited) |
| Languages | German (primary), English |
| Tokenizer | Kolibri UniBPE, 128,000 vocabulary (inherited) |
| Context length | 262,144 native; up to 1,048,576 validated on the base. Recommended ≤ 262,144 |
| Precision | FP8 (float8_e4m3fn) weights, as base |
| Reasoning effort | none, low, medium, high (inherited) |
| Augmentation mode | coach, copilot, responder (new) |
| Tool calling | Yes (inherited); optional MCP audit mode |
| Hardware | ~78 GB FP8 weights. Minimum: 2× A100 80 GB, 2× H100, 1× H200, 1× B200/B300 (as base) |
| Knowledge cutoff | Implicit knowledge as base (EN/DE 18 June 2026) |
| License | Apache 2.0 (base weights' license retained; NOTICE preserved) |
| Release date | TBA |
🏋️ Training Plan
Elster-1 changes behaviour, not architecture. All training is post-training on top of Kolibri-1.
1. Supervised fine-tuning (SFT)
- Reasoning records from LSTR. Traces produced by the LSTR harness's six-step chain, kept only when the task passed a deterministic grader (never LLM-as-judge), following the harness self-evolution policy. Provenance (task, score, supervision type) is recorded per sample.
- German-first traces. Reasoning and answers are written in German when the prompt is German, using Kolibri's language-consistency approach. No English-to-German translation of reasoning traces.
- Role corpora. Synthetic, role-specific task sets for the six roles above, generated from public-domain and permissively licensed sources (standards summaries, public guidelines, legislative texts). No customer data.
- Mode data. Paired examples of the same task in
coach,copilot, andrespondermodes.
2. Reinforcement learning (RL)
- Verifiable rewards per capability: constraint satisfaction (numeric checker), falsifiability of stated hypotheses (structural check), divergence between perspectives (embedding distance with a minimum threshold), calibration (Brier score of stated confidence vs. graded correctness).
- Abstention preserved: Merlin-Arthur style episodes are retained so the finetune does not erode the base's "I don't know" behaviour.
- Randomised harnesses and tools, as in Kolibri post-training, to avoid overfitting to one agent scaffold (including the LSTR harness itself).
- Augmentation reward (coach mode): rewards withholding the answer until the user has proposed one, and penalises solving the task for the user.
3. Compute and regulatory position
Planned post-training compute: TBA. Kolibri-1's training compute is 6.4×10²³ FLOPs. Under the European Commission's GPAI guidelines, a downstream modifier becomes a GPAI provider in its own right only if the modification uses more than one third of the original training compute (≈2.1×10²³ FLOPs here). Florentin One expects Elster-1 to stay well below this; we will confirm it with measured figures at release.
📈 Evaluation Plan
All Elster-1 columns are TBA. The base reference column is Kolibri-1's published post-training result, so a regression caused by the finetune is visible.
Guardrail metrics (must not regress vs. base)
| Benchmark | Kolibri-1 (base, published) | Elster-1 |
|---|---|---|
| Overall (DE) | 70.8 | TBA |
| GPQA Diamond (DE) | 81.3 | TBA |
| AA-Omniscience Non-Hallucination Rate | 44.0 | TBA |
| RGB Negative (Abstention) | 85.6 | TBA |
| M/A grounding score (SQuAD) | 23.4 | TBA |
| German Public Sector (Industry RAG) | 75.0 | TBA |
| IFBench (loose-prompt) | 78.1 | TBA |
Capability metrics (new, Florentin One harness)
| Capability | Metric | Grader | Elster-1 |
|---|---|---|---|
| Self-awareness | Expected calibration error on stated confidence; claim-type accuracy | Deterministic | TBA |
| Structured thinking | Step validity and dependency correctness on planning tasks | Deterministic | TBA |
| Multi-perspective | Share of answers with ≥2 non-redundant perspectives | Rule + embedding threshold | TBA |
| Hypothesis testing | Share of hypotheses with a stated falsifier and next check | Structural | TBA |
| Dialectical thinking | Antithesis quality (steelman rate) on contested cases | Rubric, human-audited sample | TBA |
| Boundary regulation | Constraint-violation rate; correct escalation rate | Deterministic checker | TBA |
| Meaning-making | Next-action relevance; zero narrative-arc violations in analytical output | Rule + human audit | TBA |
Role proxies and augmentation outcomes
| Suite | What it measures | Elster-1 |
|---|---|---|
| Six role proxies (farmer … knowledge worker) | Task success in synthetic environments per value-chain node | TBA |
| Handoff fidelity | Whether the next role can act on a handoff block without asking back | TBA |
| Unaided retention | User performance on a follow-up task without the model, coach vs. copilot vs. no AI (pilot study) |
TBA |
The last row is the one that matters for the Klarheit thesis: does Elster-1 leave people more capable? It requires a human study with consent and ethics review. We will report it even if the result is unfavourable.
🚀 Usage (planned)
Elster-1 keeps Kolibri-1's architecture, so it serves with the same stack.
pip install "aleph-alpha-inference>=1.0"
vllm serve florentin-one/Elster-1 --kv-cache-dtype fp8 \
--reasoning-parser kolibri1 \
--tool-call-parser kolibri1 \
--enable-auto-tool-choice
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="florentin-one/Elster-1",
messages=[
{"role": "system", "content": "Rolle: Disponent:in, Kunststoff-Spritzguss, 3 Linien."},
{"role": "user", "content": "Eilauftrag 4.000 Teile bis Freitag. Linie 2 hat Werkzeugwechsel Mittwoch. Wie planen?"},
],
extra_body={
"chat_template_kwargs": {
"reasoning_effort": "high",
"augmentation_mode": "copilot", # coach | copilot | responder
"emit_handoff": True,
}
},
)
print(response.choices[0].message.content)
Output contract
For gated (complex) requests, every response contains:
- Claims with type and confidence: fact / inference / speculation, 0.0–1.0. Confidence comes from the model's calibrated estimate and is never invented to sound certain.
- Reasoning summary: which capabilities were applied and the key insight from each.
- At least one boundary or failure mode with a mitigation.
- Open checks: what the human should verify before acting.
- Meaning for your next decision: one or two sentences, role-specific.
- AI disclosure (EU AI Act Art. 50): output is marked as AI-generated.
Recommended parameters
| Parameter | Value |
|---|---|
temperature |
1.0 (as base) |
top_p |
0.97 (as base) |
top_k |
128 (as base) |
reasoning_effort |
high for gated tasks, none for lookups |
augmentation_mode |
copilot default; coach for training contexts |
🎯 Intended Uses and Limitations
Primary use cases
- Fault isolation and repair planning from manuals, logs, and symptom descriptions
- Production and logistics scheduling under explicit capacity and deadline constraints
- Experiment design, hypothesis framing, and literature synthesis
- Clinical documentation, guideline lookup, and evidence summaries for clinician review
- Nutrient and field-operation planning against regulatory limits
- Cross-role synthesis and handoffs along a value chain
- Training and skill-building through
coachmode
AI system types
Elster-1 is intended for human-reviewed assistants and agentic workflows where a person approves any consequential action, as Kolibri-1's card specifies for the base. In decision-support systems it sits on the advisory side.
Limitations
- Text-only. Field photos, scans, imaging, and sensor streams need a separate perception component.
- Small active size. At 3.46B active parameters, deep multi-hop reasoning may trail much larger models; the base's HLE (DE) is 15.9. Use
reasoning_effort: highand the MCP audit mode for high-stakes analysis. - Inherited knowledge cutoff (June 2026). Regulations such as fertiliser rules or clinical guidelines change; current versions must come from retrieval.
- Calibration is trained, not guaranteed. Stated confidence can be wrong, especially outside the six role domains.
- Reasoning overhead. The full chain adds tokens and latency; the complexity gate limits but does not remove this.
- Synthetic role data. Role proxies approximate real work; deployment-specific validation is required.
Out-of-scope uses
- Diagnosis, triage decisions, or treatment recommendations without clinician decision
- Personnel scoring, task allocation by personal traits, or performance monitoring
- Direct control of machines, vehicles, or safety systems
- Predictive maintenance, HR, marketing, customer support
- Prohibited practices under EU AI Act Art. 5
- Any use that removes a human from a decision the output informs
⚖️ Ethical Considerations
Disclosure principle: we list what Elster-1 is bad at before we sell what it's good at.
- Augmentation by construction.
coachmode and the unaided-retention study exist because convenience has a cognitive cost. Success is measured by what users can do afterwards, not only by task completion. - Human as final authority. Boundary regulation escalates instead of deciding. The output contract always names what the human must check.
- Sovereignty. Local inference by default; no reasoning call leaves the deployment unless MCP audit mode is enabled (then GDPR Art. 28 applies to that endpoint).
- Transparency. Lineage is stated plainly: Elster-1 is a finetune of Kolibri-1 by Aleph Alpha, with Florentin One reasoning data. Aleph Alpha is not responsible for Elster-1.
- Bias. Inherits the base's systemic and political-bias mitigations. Role data will be audited for regional (urban vs. rural, East vs. West), gender, and professional-hierarchy bias (e.g. deferring to the clinician's or engineer's view by default). Method: TBA.
- AI literacy (Art. 4). This card and the Klarheit Manifest serve as literacy material for deploying organisations.
Regulatory map
| Area | Position |
|---|---|
| EU AI Act, GPAI | Downstream modifier; expected below the one-third compute threshold (to be confirmed) |
| EU AI Act, high-risk | Not intended for Annex I (medical devices) or Annex III (employment) uses; deployers using it there take on high-risk obligations |
| Art. 50 transparency | AI-generated output disclosed |
| GDPR | No personal data in training sets; PII refusal behaviour inherited; local processing default |
| BMFTR value-chain directive (31.07.2026) | Positioned as domain-specific along vertical chains; uses RAG, agentic workflows, XAI, reliability checks, human-in-the-loop, data provenance |
🌱 Sustainability
Base pre-training energy: ~9.5×10² MWh (Kolibri-1, estimated). Elster-1 post-training energy: TBA, measured with the same method (node power × PUE × runtime). The small active size enables single-GPU serving, which keeps inference energy per request low.
📄 License and Attribution
Elster-1 weights are released under Apache 2.0, consistent with the base model. The NOTICE and attribution for Aleph-Alpha/Kolibri-1 are preserved. Kolibri-1's license covers its weights and configuration only; Elster-1 uses no other Aleph Alpha artifacts.
@misc{florentinone2026elster1,
title = {Elster-1: Augmented Reasoning Along the German Value Chain},
author = {Florentin One Team},
year = {2026},
note = {Concept model card v0.1. Finetune of Aleph-Alpha/Kolibri-1.}
}
Point of contact (incl. rightsholders, GPAI Code of Practice Measure 1.5): compliance@florentin-one.de
Florentin One · Hannover, Germany. Under EU AI Act Art. 50, this card is identified as an AI-assisted artifact.
- Downloads last month
- -