LLaDA 1.5
The first preference-aligned diffusion language model: LLaDA-8B tuned with variance-reduced preference optimization (VRPO), MIT licensed.
About
LLaDA 1.5 is the GSAI team's follow-up to LLaDA-8B, the diffusion language model that showed text can be generated by iterative denoising instead of left-to-right autoregression. Where the base LLaDA was pretrained with a masked-diffusion objective, 1.5 adds a preference-alignment stage built on VRPO, a variance-reduced preference optimization method designed for the diffusion setting. The team released it in late May 2025 with weights, demo, and eval code, and the Hugging Face repo has since passed 15,000 downloads.
The release matters less for raw scores than for what it proved: alignment techniques from the autoregressive world can be ported to diffusion LMs. GSAI-ML has kept building on the line since, shipping LLaDA-V (a diffusion vision-language model), the LLaDA-MoE diffusion mixture-of-experts, and iLLaDA in 2026. The whole family is MIT licensed, which makes 1.5 one of the few diffusion LMs you can actually build on commercially. Treat its benchmark deltas as team-reported; diffusion models are still hard to compare against autoregressive leaderboards apples-to-apples.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- GSAI-ML reports that VRPO preference alignment lifts LLaDA 1.5 over the base LLaDA-8B on general capability benchmarks (provider claim, LLaDA 1.5 demo page)
- No independent benchmark table verified; diffusion-LM evals are not directly comparable to autoregressive leaderboards
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: GSAI-ML announced LLaDA 1.5 on May 25, 2025 (team GitHub news log); HF repo GSAI-ML/LLaDA-1.5 created May 28, 2025 with 15,451 downloads and 41 likes (HF API, Oct 2026); MIT license on the model card. Open weights. Strengths: first preference-aligned diffusion LM, permissive MIT license, active research line (LLaDA-V, LLaDA-MoE, iLLaDA followed). Docked for: (1) benchmark deltas are team-reported with no independent replication; (2) diffusion LMs remain a research niche with thin tooling and eval comparability. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · HF repo GSAI-ML/LLaDA-1.5 created 2025-05-28; 15,451 downloads, 41 likes (HF API Oct 2026)
- · Announced by GSAI-ML team May 25, 2025 with VRPO preference alignment (team GitHub news log)
- · Active follow-through: LLaDA-V, LLaDA-MoE, iLLaDA released after
- · Benchmark deltas are team-reported; diffusion LM evals not comparable to autoregressive leaderboards