alibaba-multimodal-industrial-ai/IndustryLLM

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

IndustryLLM is a 35.1 billion parameter Mixture-of-Experts (MoE) industrial language model developed by the Multimodal and Industrial AI Team at Alibaba, based on Qwen3.5-35B-A3B-Base. It is specifically designed for industrial procurement, integrating deep engineering knowledge from a 100B-token industrial corpus including national standards and B2B transaction records. The model excels at robust procurement query structuring, resolving colloquial jargon and conflicting specifications, and features both a Reasoning-Enabled ('Think') mode for complex analysis and a Direct-Response ('No-Think') mode for low-latency production environments.

Loading preview...

IndustryLLM: Failure-Driven LLM for Industrial Procurement

IndustryLLM, developed by the Multimodal and Industrial AI Team at Alibaba, is a 35.1 billion parameter Mixture-of-Experts (MoE) model built upon Qwen3.5-35B-A3B-Base. It features approximately 3 billion parameters activated per token, combining industrial engineering knowledge with efficient inference.

Key Capabilities

  • Authoritative Engineering & Standards Grounding: Pre-trained on a curated ≈100B-token industrial corpus, including 5B tokens of national standards (e.g., GB/T) and 10B tokens of authentic enterprise B2B transaction records.
  • Failure-Driven Data Reconstruction: Addresses factual brittleness and register mismatch through multi-register rewriting, confidence-routed factual editing, and error-targeted QA synthesis.
  • Robust Procurement Query Structuring: Interprets colloquial buyer jargon, phonetic typos (e.g., 42络钼 → 42CrMo), and truncated standard codes (16674 → GB/T 16674), while identifying contradictory specifications.
  • Dual Operating Modes: Offers a Reasoning-Enabled Mode (Think) for deep multi-step engineering analysis and a Direct-Response Mode (No-Think) for latency-critical production environments requiring sub-2s response times.
  • Text-Only Focus: While inheriting multimodal architecture, its training specifically adapted language model parameters, keeping vision modules frozen to isolate text-based engineering reasoning.

Good For

  • Industrial Procurement: Interpreting and structuring complex, often ambiguous, buyer inquiries for industrial products.
  • Supply Chain Optimization: Automating the initial stages of product sourcing by standardizing specifications and identifying potential conflicts.
  • Engineering Analysis: Leveraging its 'Think' mode for detailed, multi-step engineering reasoning tasks within industrial contexts.
  • Applications Requiring Low Latency: Utilizing the 'No-Think' mode for rapid query processing in production systems.

Finetunes
1 models