Models
Browse foundation and domain models available across Discover. Find deployment guides.
Qwen3.8-27B on RTX PRO 6000: NVFP4, MTP, and DSpark with SGLang
Aug 16, 2026 · Discover
A hands-on follow-up to the Qwen3.6 vLLM and Qwen3.8 RTX 5090 guides: serve Qwen3.8-27B on a 96 GB RTX PRO 6000 with SGLang — NVFP4+MTP, official FP8 with explicit FP8 KV, native 262K context, real multi-request overlap, and radix-cache checks.
Qwen3.8-27B on a Single RTX 5090: Full-Context SGLang Deployment
Aug 15, 2026 · Discover
A hands-on guide to serving the dense 27B hybrid-GDN vision-language model Qwen3.8-27B-NVFP4 with SGLang on a single 32 GB RTX 5090 — covering the GDN state/KV memory split, the --mamba-full-memory-ratio knob, pushing context to 200K tokens, benchmarked throughput at 1/2/4 concurrency, and why DSpark speculative decoding does not fit this card.
Hosting Ornith-1.0-35B-FP8 for Field Coding Agents
Jun 27, 2026 · Discover
A guide to self-hosting Ornith-1.0-35B-FP8 with vLLM for vibe coding, field deployed engineers, Claude Code compatibility, SSH tunnels, and practical performance checks.
Input Security Review Model
Jun 24, 2026 · Discover
A small review model that inspects user prompts, documents, and AI tool requests before they reach downstream LLMs or agents — classifying sensitive data, intent, prompt-injection risk, and unsafe tool use, then returning a structured handling recommendation. Built on Qwen3 0.6B with MLX-LM and CUDA/Swift LoRA experiments over a 7-batch synthetic dataset.
Training Qwen3 Reranker with ms-swift
Jun 23, 2026 · Discover
End-to-end practical tutorial for fine-tuning Qwen3-Reranker-0.6B on custom MTEB-style retrieval data using ms-swift generative_reranker task, pointwise vs listwise loss, LoRA rank experiments, 3-epoch "train longer" runs, custom holdout evaluation (MRR, Success@K, NDCG, avg rank), and measured gains from each change.
Deploying Qwen3 Embedding with vLLM
Jun 22, 2026 · Discover
A guide to serving Qwen3-Embedding on a Linux GPU server using vLLM, with an OpenAI-compatible HTTP API.
Deploying Qwen3 Reranker with vLLM
Jun 22, 2026 · Discover
A guide to serving Qwen3-Reranker on a Linux GPU server using vLLM, with score and rerank HTTP APIs.
Deploying Qwen3.6-27B-FP8 with vLLM
Jun 22, 2026 · Discover
A production-oriented guide to serving Qwen3.6-27B-FP8 on a single NVIDIA GPU with vLLM — covering memory tuning, MTP speculative decoding, OpenAI API compatibility, and benchmarked throughput.