Dimensions overall score 8.0

• InCoder-32B is the first 32B-parameter code foundation model unifying code intelligence across chip design, GPU kernel optimization, embedded systems, compiler optimization, and 3D modeling.technical 90%
• The model is trained from scratch using a multi-stage process: general code pre-training, curated industrial code annealing, mid-training extending context from 8K to 128K tokens, and post-training with execution-grounded verification.method 90%
• InCoder-32B achieves highly competitive performance on general code tasks.result 85%
• InCoder-32B establishes strong open-source baselines across industrial domains.result 85%
• Current code large language models' performance degrades significantly in industrial scenarios requiring reasoning about hardware semantics, specialized language constructs, and strict resource constraints.limitation 90%
• The model's training involves the use of synthetic industrial reasoning data.method 90%
• The evaluation of InCoder-32B was conducted on 14 mainstream general code benchmarks and 9 industrial benchmarks spanning 4 specialized domains.result 90%
• A key product opportunity for this model is automating complex, specialized coding tasks for semiconductor companies and GPU manufacturers.market 80%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card