Primary use stated or inferred cautiously from official documentation.
M
MICROSOFT / microsoft/AesCode-32B
AesCode 32B
Microsoft's multimodal family turns content requirements and optional visual references into editable HTML/CSS documents. The 8B and 32B checkpoints derive from Qwen3-VL-8B-Instruct and Qwen3-VL-32B-Instruct respectively. They produce code, not raster images or video files. The reference provides composition and style, while the prompt remains the source of truth for text, figures and relationships. The verified public release is October 11, 2026; earlier repository creation is not its release date.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | microsoft/AesCode-32B |
|---|---|
| Model type | Code · Multimodal |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosting with Transformers or vLLM |
| Release | 11 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image |
| Exit | Text · Code |
| Context window | 24,576 tokens in the published evaluation |
| maximum output | 12,000 tokens in the published evaluation |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Complete HTML document with CSS and ECharts |
| Batch processing | Batch inference with self-hosted vLLM |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · Transformers · vLLM |
Slides, posters, dashboards and reports with verifiable tables and charts. The 32B variant uses SFT and 520 GDPO steps with seven reward channels. The card estimates roughly 65 GB of accelerator memory for bf16 parameters alone, before cache and activations. It reports better aggregate results than 8B in its experiment, not universal superiority. It declares English and requires Transformers 4.57 or later.
The 24,576-token context and 12,000-token output belong to the published experiment. config.json declares 262,144 positions, but this does not establish evaluated quality at that limit. Likewise, two images allowed in the server example are not a universal guarantee. Generated HTML can contain scripts: review it and render it in an isolated environment without credentials or network access before publication. Do not confuse AesCode with the older AesCoder-4B or the AesCode-358K dataset.
03 / CAPABILITIES
What it can do
Orientation
Slides, posters, dashboards and reports with verifiable tables and charts. The 32B variant uses SFT and 520 GDPO steps with seven reward channels. The card estimates roughly 65 GB of accelerator memory for bf16 parameters alone, before cache and activations. It reports better aggregate results than 8B in its experiment, not universal superiority. It declares English and requires Transformers 4.57 or later.
Context
24,576 tokens in the published evaluation · 12,000 tokens in the published evaluation
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Input modalities
TextImage
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
The weights declare Apache 2.0; the official repository's code uses MIT. These are different licenses for different artifacts. We have not verified a hosted commercial API or token prices: numeric fields remain empty. Budget for GPU, memory, rendering and operations. The reference-creation pipeline may use external image and vision providers whose costs and terms are not included in the checkpoint.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Microsoft reports nine metrics on 300 infographics, averaging three generations per prompt without selecting the best. Rule combines Text, Bound and Chart; Visual combines Content, Layout and Style; Overall averages Rule and Visual. Recorded results use a visual reference, temperature 0.8 and top-p 0.95. They are not a general coding benchmark or independent Inferama tests. Full reproduction requires equivalent data and judges: the repository publishes code, not datasets.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| AesCode Infographics Text | 95,34 % | Percentage | View source ↗ |
| AesCode Infographics Bound | 97,27 % | Percentage | View source ↗ |
| AesCode Infographics Chart | 90,37 % | Percentage | View source ↗ |
| AesCode Infographics Rule | 94,33 % | Percentage | View source ↗ |
| AesCode Infographics Content | 85,76 % | Percentage | View source ↗ |
| AesCode Infographics Layout | 90,58 % | Percentage | View source ↗ |
| AesCode Infographics Style | 55,99 % | Percentage | View source ↗ |
| AesCode Infographics Visual | 77,44 % | Percentage | View source ↗ |
| AesCode Infographics Overall | 85,89 % | Percentage | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
AesCode 32B
Verified public checkpoint release. Repository created September 29; this is not used as its release date.
ANALYSIS
Related analysis
AesCode 32B: Verification Protocol for Local and Self-Hosted Deployment Costs
An operational guide to auditing the real costs of AesCode 32B, distinguishing between technical specifications confirmed by Microsoft and economic variables that require internal estimation, given that no public API pricing exists for this model.
18 Oct 2026 ↗
AesCode Infographics Content: What It Measures, How It Runs, and Why Direct Comparison Is Blocked
A methodological audit of the 'Content' metric within the AesCode benchmark. This analysis defines its role in evaluating semantic fidelity of code-generated infographics, its integration with other visual metrics, and the specific limitations that prevent direct comparison with other industry standards.
17 Oct 2026 ↗
AesCode Infographics Chart: what the evidence lets us verify—and what is missing for comparison
Public documentation describes the Chart channel, reports results, and provides evaluation code. But the datasets are not published, limiting independent reproduction and interpretation of the scores.
16 Oct 2026 ↗
AesCode Infographics Bound: What It Measures, How It Runs, and What It Cannot Compare
A methodological audit of the 'Bound' metric within Microsoft's AesCode evaluation suite. This analysis examines its role as a programmatic constraint in infographic generation, its integration with other metrics, and the limitations it imposes on cross-model comparability.
16 Oct 2026 ↗07 / SOURCES