Ornith-1.5-35B-A3B matters because it pushes self-improvement beyond fixed human-curated tasks: the model continuously generates new tasks, designs scaffolds, and learns solution rollouts via reinforcement learning, which drives better search trajectories and stronger agentic coding behavior than similarly sized dense models.
Key Capabilities
- Self-improving agentic training loop: the model is trained to propose tasks, build solution scaffolds, and produce rollouts that are jointly optimized by RL — this reduces reliance on static human-written harnesses and can discover higher-yield strategies autonomously.
- MoE efficiency for long-context agenting: a ~35B MoE that activates ≈3B parameters per token, enabling large-capacity behavior while keeping per-request compute comparable to smaller dense models; natively supports 262,144-token context windows and validated extensions (YaRN) toward ~1M tokens.
- Agent & tool-first output design: emits explicit reasoning blocks (chain-of-thought) and well-formed tool-call function blocks compatible with OpenAI-style tool APIs, making it suitable for tool-enabled agent frameworks and terminal coding agents.
- Production-friendly serving: provided recipes for vLLM/SGLang, GGUF builds for local inference (llama.cpp/Ollama), and recommended sampling settings for reproducible benchmark runs.
Who Should Use It and Tradeoffs
Great fit if you need a model focused on autonomous coding agents and tool-enabled workflows, want very large context windows for multi-file/codebase reasoning, and can allocate multi-GPU serving (or use GGUF for local inference). Look elsewhere if you need a tiny single-GPU dense model for low-resource devices, if deterministic short-answer tasks are primary, or if you cannot accept the infrastructure cost of MoE serving (recommended ~2×80GB GPUs for full bf16 serving with large context).
