agentic coding

Bilingual

BTL-3: A 27B open-weight agent model for agentic coding and structural tool use
2 months ago
- BTL-3 is a 27B open-weight agent model designed for agentic coding, structural tool use, repository work, failure recovery, and long multi-turn execution.
- It achieves high scores on benchmarks: 95.1% HumanEval pass@1, 88.5% BFCL v4 AST, and 88.1% LiveCodeBench v6.
- A Compact edition (8.39 GB) packages the full model into a single file with 92.2% behavior retention, enabling native local inference.
- The model is a PEFT LoRA adapter (rank 32, alpha 64) based on Qwen3.6-27B, with a 262,144-token context window and Apache-2.0 license.
- It can be used with Transformers (via PEFT) or vLLM, with recommended non-thinking mode for stability.
- Intended use cases include coding, debugging, test-driven repair, structured function calling, and long multi-turn tasks requiring verification and recovery.
- Operational guidance recommends sandboxing generated code and requiring confirmation for high-impact actions.
Laguna S 2.1
2 months ago
- Laguna S 2.1 is a 118B-parameter Mixture-of-Experts model with 8B active parameters per token and a 1M-token context window, developed in under nine weeks.
- It achieves competitive scores against much larger models on long-horizon coding benchmarks, scoring 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE v1.1.
- All evaluation trajectories are publicly released at trajectories.poolside.ai for transparency and reproducibility.
- Case studies show the model building a browser engine from scratch, optimizing its own harness (5.2% speedup, 70% memory reduction), and rediscovering a mathematical proof.
- The model offers two thinking modes: off and max, with max significantly improving performance (e.g., Terminal-Bench 2.1 from 60.4% to 70.2%).
- Known limitations include harness overfitting, nested tool call issues, and longer than expected thinking duration on difficult problems.
- Key training improvements include larger rollout budgets, better sandbox infrastructure, multi-harness rollouts, and RL in FP8 precision.
- The model focuses on agentic coding capabilities and on recovering hidden reasoning through reinforcement learning.
- Laguna S 2.1 is available via multiple platforms (Hugging Face, OpenRouter, Baseten) with open weights under OpenMDW-1.1.
Experiences with Local Models for Coding
2 months ago
- Small local models for agentic coding were evaluated using a viability funnel covering RAM, speed, tool calling, functional correctness, conversation handling, task complexity, and code quality.
- The evaluation process went through manual evals, automated evals, and day-to-day use, with tasks including sorting/cumulating a bar chart and creating a bar chart from access log data in JavaScript/TypeScript, plus Bash and Python scripts.
- Results varied: some models like Qwen Coder Next 80B MoE produced functional code but crashed in extended conversations, while others like Gemma 4 26B succeeded manually but failed in automated tests, with performance differing between machines.
- Day-to-day use with Qwen3.6 35B MoE showed success for small, well-defined tasks (e.g., Bash/Python scripts, website updates) but struggled with complex logic, requiring careful task selection and more code review.
- Key factors affecting viability include task complexity, number of files to edit, specificity of instructions, and tech stack, with local models offering a 'detox' by encouraging slower, more thoughtful coding but lagging behind larger models in autonomy.

agentic coding

BTL-3: A 27B open-weight agent model for agentic coding and structural tool use

Laguna S 2.1

Experiences with Local Models for Coding