Project | Code | Collection
# OpenThinkerAgent-8B-RL **OpenThoughts-Agent** is an open-source effort to curate the best datasets for training agents. Our release includes [datasets](https://huggingface.co/collections/open-thoughts/openthinker-agent), [models](https://huggingface.co/collections/open-thoughts/openthinker-agent) and our [research codebase](https://github.com/open-thoughts/OpenThoughts-Agent). [OpenThinkerAgent-8B-RL](https://huggingface.co/open-thoughts/OpenThinkerAgent-8B-RL) is the **final, RL-trained** 8B agentic checkpoint of the OpenThoughts-Agent SFT→RL recipe. Starting from the cold-start SFT base [OpenThinkerAgent-8B-ColdStartSFTForRL](https://huggingface.co/open-thoughts/OpenThinkerAgent-8B-ColdStartSFTForRL), it is further trained with on-policy reinforcement learning on the [OpenThoughts-Agent-RL-5K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-RL-5K) task set. This checkpoint corresponds to **RL step 45**. > **Architecture note.** Although the upstream lineage carries a `GLM-4.7` label (which refers to the *teacher* used for the cold-start SFT trajectories, not the student), this model is a **Qwen3-8B**. Its `config.json` reports `model_type: qwen3`, `architectures: ["Qwen3ForCausalLM"]`, 36 layers, hidden size 4096, 32 attention heads / 8 KV heads, and a 40,960-token context — i.e. standard Qwen3-8B. - **Homepage:** https://www.openthoughts.ai/blog/agent - **Repository:** https://github.com/open-thoughts/OpenThoughts-Agent # Model details - **Base (pre-RL) model:** [OpenThinkerAgent-8B-ColdStartSFTForRL](https://huggingface.co/open-thoughts/OpenThinkerAgent-8B-ColdStartSFTForRL) (itself an SFT of [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)) - **Architecture:** Qwen3 (`Qwen3ForCausalLM`), 36 layers, hidden size 4096, 32 attention heads, 8 KV heads, RoPE θ = 1e6 - **Context length:** 40,960 tokens (max position embeddings); RL rollouts used a 32,768-token serving window - **Vocabulary:** 151,936 tokens - **Precision:** bf16 - **Checkpoint:** RL step 45 # The SFT → RL recipe 1. [OpenThoughts-Agent-SFT-ColdStartForRL-10K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-SFT-ColdStartForRL-10K) — cold-start SFT trajectories. 2. [OpenThinkerAgent-8B-ColdStartSFTForRL](https://huggingface.co/open-thoughts/OpenThinkerAgent-8B-ColdStartSFTForRL) — Qwen3-8B after cold-start SFT (the pre-RL base). 3. [OpenThoughts-Agent-RL-5K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-RL-5K) — the 5,000 on-policy RL tasks. 4. **[OpenThinkerAgent-8B-RL](https://huggingface.co/open-thoughts/OpenThinkerAgent-8B-RL)** — this model, the final RL'd checkpoint (step 45). # Training data - **Cold-start SFT:** [OpenThoughts-Agent-SFT-ColdStartForRL-10K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-SFT-ColdStartForRL-10K) (9,437 task/trajectory pairs). - **RL tasks:** [OpenThoughts-Agent-RL-5K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-RL-5K) (5,000 `pymethods2test-large` tasks); the policy rolls out against each task in a Daytona sandbox and is rewarded by the task's test verifier. # Training procedure On-policy RL with the OpenThoughts-Agent codebase (SkyRL), recorded in the run config shipped with this repo (`swesmith-fixthink-pymethods2test_rl_config.json`): - **Algorithm:** RLOO-n advantage estimator (`advantage_estimator=rloo_n`), no KL loss (`use_kl_loss=false`, `kl_loss_coef=0.0`) - **PPO clip:** eps_clip_low/high = 0.2, loss reduction = token_mean - **Optimizer:** AdamW, learning_rate 5e-6, weight_decay 0.0, betas (0.9, 0.999) - **Batch:** train_batch_size 64, policy_mini_batch_size 64 - **Rollouts:** vLLM backend, 8 samples per prompt, sampling temperature 0.7 / top_p 0.95 / top_k 20, max generate length 4096, served at 32,768-token context - **Harness:** terminus-2 agent in Daytona sandboxes; interleaved thinking enabled - **Strategy:** FSDP2; HF checkpoint exported every 5 RL steps; **this artifact is step 45** # Intended uses & limitations This is an **agentic coding model**: it is designed to operate as a tool-using agent in the terminus-2 harness (issuing shell commands / edits and reasoning over terminal output) to solve software-engineering tasks. It inherits Qwen3-8B's general capabilities plus agentic behaviour from cold-start SFT and the RL stage. Limitations: outputs (including shell commands) may be incorrect or unsafe and should be executed only in sandboxed environments with review; the RL stage optimized for the `pymethods2test`/SWE-Smith-style task distribution and may generalize unevenly to other domains. > **Evaluation:** No verified agentic-benchmark numbers are published for this specific 8B RL checkpoint in the source artifact; evaluation results are **TBD**. (The flagship [OpenThinkerAgent-32B](https://huggingface.co/open-thoughts/OpenThinkerAgent-32B) card reports the project's benchmark suite for the 32B SFT line.) # Links - 🌐 [OpenThoughts-Agent project page](https://www.openthoughts.ai/blog/agent) - 💻 [OpenThoughts-Agent GitHub repository](https://github.com/open-thoughts/OpenThoughts-Agent) - 📚 [OpenThinker-Agent collection](https://huggingface.co/collections/open-thoughts/openthinker-agent) - 🤖 [Pre-RL base model: OpenThinkerAgent-8B-ColdStartSFTForRL](https://huggingface.co/open-thoughts/OpenThinkerAgent-8B-ColdStartSFTForRL) - 🧠 [RL tasks: OpenThoughts-Agent-RL-5K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-RL-5K) - 🧠 [Cold-start SFT dataset: OpenThoughts-Agent-SFT-ColdStartForRL-10K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-SFT-ColdStartForRL-10K) # Citation ``` @misc{openthoughts-agent, author = {Team, OpenThoughts-Agent}, title = {{OpenThoughts-Agent: Data Recipes for Agentic Models}}, howpublished = {https://www.openthoughts.ai/blog/agent}, year = {2026} } ```