Model reference · open weights
LaSER-Qwen3 is an open-weight embedding model from Alibaba-NLP, listed in the AxForge catalogue. AxForge can bring it up on EU-owned hardware for you on request — with the licence handled where one is required.
About
LaSER-Qwen3-0.6B LaSER (Latent Space Explicit Reasoning) is a self-distillation framework that internalizes explicit Chain-of-Thought reasoning into the latent space of dense retrievers, enabling the model to "think silently" through continuous latent tokens. LaSER-Qwen3-0.6B is a 0.6B-parameter dense retriever built on Qwen/Qwen3-0.6B. 📄 Paper: LaSER: Internalizing Explicit Reasoning into Latent Space for Dense Retrieval 💻 Code: https://github.com/ignorejjj/LaSER Model Summary How It Works Unlike standard dense retrievers that encode queries in a single forward pass, LaSER generates K continuous latent thinking tokens autoregressively in the embedding space: 1. Encode the input text into embeddings 2. At each thinking step, project the last hidden state through the LM head → softmax → compute a probability-weighted soft token from the embedding table 3. Append the soft token and repeat for K steps (using KV caching for efficiency) 4. Mean-pool the hidden states from all K thinking steps → L2 normalize This enables complex reasoning while maintaining the inference efficiency of standard dense retrievers (~1.7× latency overhead, only ~0.3% of rewrite-then-retrieve pipelines). Usage Direct Usage with Transformers Batch Encoding Evaluation Results BRIGHT Benchmark (nDCG@10) — In-Domain FollowIR & BrowseComp-Plus — Out-of-Domain † indicates methods that use an external LLM to rewrite queries during inference. Training Details - Training Data: 81K query-document pairs from ReasonEmb, each with a CoT reasoning path generated by GPT-4o-mini - Method: LoRA fine-tuning (r=64, α=32) for 1 epoch on 4×A100 GPUs - Loss: Contrastive learning + Output-level KL distillation (λ₂=10) + Process-level trajectory alignment (λ₃=0.1) - Temperature: τ=0.02 - Thinking Steps: K=3 Model Family Citation
Summarised from the published model card. Read the full card on the HuggingFace links below.
Specifications
| Maker | Alibaba-NLP |
|---|---|
| Type | Embedding models |
| Parameters (lead) | 596M |
| Context | 40k tokens |
| Variants | 3 |
| Runs with | transformers |
| Based on | Qwen/Qwen3-0.6B |
| Released | 2026-03-31 |
| Popularity | 693 downloads / month |
| Likes | 6 |
| Licence | Open weights |
How it works
Variants
Open weights ship in several sizes and precisions. One page, all the variants — pick the one that fits your GPU. VRAM figures are estimates from model size.
Using it via the API
Once AxForge deploys laser-qwen3 for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (laser-qwen3 below is illustrative; you get the exact model name on deployment.)
$ curl -sS https://api.axforge.ai/v1/embeddings \
-H "Authorization: Bearer $AXFORGE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"laser-qwen3","input":"text to embed"}'
Licence
Open weights under mit — commercial use is permitted. Deploy it on AxForge EU hardware on request. Read the licence ↗
Explore