Qwen3-8B

Type: AI model

Qwen3-8B is an 8-billion-parameter open-weight model in Alibaba’s Qwen3 family, licensed under Apache 2.0 and distributed on Hugging Face. It supports switching between a reasoning ‘thinking’ mode and an efficient non-thinking mode within one model. GROUNDING tracks Qwen3-8B’s benchmarks, deployment, and use in local and self-hosted setups.

Recent Updates

• 2026-07-03: Mechanistic analysis of authority bias in LLM sycophancy (cs.CL updates on arXiv.org) · arxiv.org — Llama-3.1-8B · Gemma-2-9B
• 2026-07-03: Procedural Memory Distillation turns rollout history into training-time supervision for self-improving language models (cs.AI updates on arXiv.org) · arxiv.org — OLMo3-Instruct-7B SDPO SCIKNOWEVAL LIVECODEBENCH
• 2026-07-03: MedAgentBench v3 exposes barriers to RL in clinical agent tasks (cs.AI updates on arXiv.org) · arxiv.org — Agents LLM Evals
• 2026-07-03: Spec-AUF trains block drafters with accept-until-fail loss support (cs.AI updates on arXiv.org) · arxiv.org
• 2026-07-03: ReContext proposes training-free evidence replay for long-context reasoning (cs.AI updates on arXiv.org) · arxiv.org — Long Context Context Engineering Qwen3-4B · Llama3-8B
• 2026-07-06: ReContext adds a training-free replay step for long-context reasoning (DAIR.AI) · arxiv.org — Long Context Context Engineering DAIR.AI · Qwen3-4B · Llama3-8B
• 2026-07-07: HASE claims to co-evolve task solutions and the evaluation harness in one agentic loop (cs.AI updates on arXiv.org) · arxiv.org — Agents LLM Evals GPT-OSS 120B
• 2026-07-07: RiskAverseOOD measures whether low-stakes risk aversion transfers to extreme stakes in language models (cs.LG updates on arXiv.org) · arxiv.org — LLM Evals Qwen3-1.7B · Qwen3-14B gemma-3-12b-it · Llama 3.1 8B Instruct
• 2026-07-08: KVpop proposes learned KV-cache eviction with future-attention supervision (cs.AI updates on arXiv.org) · arxiv.org — arXiv Lukas Hauzenberger · Qwen3-4B
• 2026-07-08: Nemotron-Labs-Diffusion proposes a tri-mode language model for AR, diffusion, and self-speculation decoding (cs.CL updates on arXiv.org) · arxiv.org — Nemotron-Labs-Diffusion Nemotron-Labs-Diffusion-3B Nemotron-Labs-Diffusion-8B Nemotron-Labs-Diffusion-14B
• 2026-07-10: DominoTree proposes training-free tree drafting for speculative decoding (cs.CL updates on arXiv.org) · arxiv.org — LLM Evals Qwen3-4B
• 2026-07-10: Jet-Long proposes a zero-shot long-context extension with dynamic bifocal RoPE (cs.LG updates on arXiv.org) · arxiv.org — Long Context Context Engineering RAG Code Agents Agents arXiv · Qwen3-1.7B · Qwen3-4B · Jet-Nemotron
• 2026-07-10: Selective Left-Shift turns test feedback into training data for low-resource code generation (cs.LG updates on arXiv.org) · arxiv.org — LLM Evals
• 2026-07-11: Jet-Long proposes dynamic zero-shot long-context extension with bifocal RoPE (cs.AI updates on arXiv.org) · arxiv.org — Long Context Context Engineering RAG Codebase Indexing Agents Qwen3-1.7B · Qwen3-4B · Jet-Nemotron
• 2026-07-14: SPARK profiles latent reasoning states and steers under-activated examples (cs.AI updates on arXiv.org) · arxiv.org — LLM Evals Qwen3 series Qwen3-4B
• 2026-07-14: SETA proposes verifiable RL environments for terminal agents (cs.AI updates on arXiv.org) · arxiv.org — Agents LLM Evals arXiv · DeepSeek-V4-Flash
• 2026-07-14: Amplitude-Only FFN Intervention for Tool-Structured LLM Inference (cs.CL updates on arXiv.org) · arxiv.org — Agents Tool Use LLM Evals Qwen3.5-9B · Qwen2.5-7B
• 2026-07-14: MET proposes theory-grounded multilingual moral reasoning with a new benchmark (cs.CL updates on arXiv.org) · arxiv.org — LLM Evals Qwen3-4B · Gemma3-4B
• 2026-07-14: Probing LLM internal states to detect confident hallucinations in financial QA (cs.CL updates on arXiv.org) · arxiv.org — LLM Evals Llama-3.1-8B · Gemma-2-9B
• 2026-07-15: ISE: Execution-grounded synthesis for multi-turn OS-agent training trajectories (cs.CL updates on arXiv.org) · arxiv.org — Agents Tool Use LLM Evals Siyuan Luo GPT-4o · Qwen3-32B mpnet-base-v2
• 2026-07-15: Function-Aware Fill-in-the-Middle Mid-Training for Coding Agent Foundation Models (cs.CL updates on arXiv.org) · arxiv.org — Code Agents Agents Tool Use LLM Evals Open Source LLMs GitHub Qwen2.5-Coder-Instruct
• 2026-07-16: Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models (cs.AI updates on arXiv.org) · arxiv.org — Agents Code Agents Tool Use GitHub Qwen2.5-Coder-Instruct
• 2026-07-20: Better Starts, Better Ends: Bootstrapped Iterative Self-Reasoning Distillation for Compressed Reasoning (cs.CL updates on arXiv.org) · arxiv.org — LLM Evals Qwen3
• 2026-07-21: Study finds answer pre-commitment in Qwen3-8B on a minimal reasoning probe (cs.CL updates on arXiv.org) · arxiv.org — LLM Evals
• 2026-07-22: Reproduction report challenges RLSD’s headline accuracy claim (‌alphaXiv) · github.com — LLM Evals alphaXiv Tinker API EasyVideoR1 veRL · Ray · vLLM · Qwen-VL · marimo wandb · Qwen3.5 4B

FAQ

What is Qwen3-8B?

Qwen3-8B is an 8-billion-parameter open-weight model in Alibaba’s Qwen3 family, licensed under Apache 2.0 and distributed on Hugging Face. It supports switching between a reasoning ‘thinking’ mode and an efficient non-thinking mode within one model. GROUNDING tracks Qwen3-8B’s benchmarks, deployment, and use in local and self-hosted setups.

What does this page track?

Dated radar mentions, source links, related concepts, and builder-relevant context for Qwen3-8B, collected automatically by GROUNDING.

When was Qwen3-8B last mentioned?

Qwen3-8B was most recently mentioned in a radar update dated 2026-07-22.

Category: Text / Language Models