Source availability
A multi-core product depends on the availability and performance characteristics of its underlying intelligence sources.
Quantel AI Research · Version 1.0 · 18 August 2026
Multi-Core Language Models
A new class of AI models for synthesized intelligence with Single-Stream Inference.
Abstract
One persistent identity for synthesized intelligence.
Modern language-model development often treats a foundation model as a single capability source. Multi-agent and routing systems can combine independently trained models, but they usually expose that collaboration as orchestration: agents, routes, subtasks, intermediate responses and aggregation.
The Multi-Core Language Model (MCLM) is a different abstraction. Specialized capabilities derived from heterogeneous foundation models are synthesized behind a persistent model identity and exposed through one inference interface. From the developer's perspective, Quantum Ultra 1 behaves as one model: one model identifier, one API request, one continuous streaming response and one consistent behavioral boundary.
Core thesisMultiple specialized intelligence sources. One persistent model identity. One inference boundary. One continuous response.
Quantel Synthesis Architecture · QSA
Capability specialization becomes a unified inference process.
QSA is the proprietary model-development and inference architecture used to construct Quantum's unified behavior. The diagram describes the logical boundary, not the complete implementation.
01Intelligence coresReasoning · coding · planning · generation
↓02Quantel Synthesis ArchitectureProprietary synthesis and inference layer
↓03Single-Stream InferenceOne model ID · one SSE token stream
What remains proprietary
Quantel does not publicly disclose internal transformations, scheduling strategies, information-transfer mechanisms, intermediate representations, synthesis policies, caching strategies, thresholds, optimization methods or infrastructure-specific mechanisms.
Single-Stream Inference · SSI
Conventional client behavior, synthesized model behavior.
SSI is the inference mode in which multiple intelligence sources contribute to one contiguous output stream without requiring visible agent coordination, external aggregation or multiple client-side inference requests.
What the client sees
Applications integrate with a stable public model ID such as quantum-ultra. Coding agents, IDE integrations and terminal assistants receive the same familiar streaming boundary.
RequestPOST /v1/responsesOne request to the Quantel gateway.
Modelquantum-ultraOne persistent product identity.
OutputContinuous SSE responseOne stream for the developer client.
Evaluation methodology
Interactive performance is part of the claim.
Quantel's latest internal cycle included more than 2,000 individual test runs over approximately one month, including coding-oriented workloads, terminal agents, repeated consistency tests, latency measurements and streaming behavior.
Quantum Ultra 1>92.7%Internal Terminal-Bench evaluation
Raw baseline~83%Same internal evaluation framework
Difference+9.7 pp≈ +11.7% relative improvement
2,000+ test runs≈ one month of evaluation50–100 output tokens/sec* in production-oriented testing
*Throughput depends on request complexity, context size, provider conditions, workload and serving configuration. These are Quantel internal results and should not be read as an independently audited or official benchmark-maintainer leaderboard score unless separately reproduced.
Transparency and limitations
Clear about the architecture's boundary.
Quantum Ultra 1 is a commercial proprietary model. Quantel distinguishes scientific disclosure from implementation disclosure and does not present the model as a foundation model trained entirely from scratch by Quantel.
New measurement questions
Traditional benchmarks may not fully characterize a composed-intelligence architecture or its latency/cost trade-offs.
Closed commercial system
Behavioral claims can be evaluated, while the complete production mechanism remains undisclosed.
Current technology attribution
The current Quantum Ultra generation incorporates technologies from DeepSeek, GLM and Kimi model families. Quantel's contribution is its architecture, synthesis methodology, inference behavior, optimization, deployment system, model identity and product boundary. Inclusion does not imply endorsement or partnership.
Continue exploring
Read the model catalog or join early access.
The research abstraction is paired with stable public model IDs and three native protocol interfaces.