Tiel-Coder 35B A3B

Tiel-Coder is a community agentic coding model built on Ornith-1.5-35B-A3B. Rechecked 2026-08-25 on the GGUF model card: it is a Mixture-of-Experts model with roughly 35B total / 3B active parameters, re-quantized with a custom imatrix and the Sharp chat template, carrying ~22 GB at 4-bit quant, licensed MIT. The card reports 12 of 25 on SWE-bench-Live, level with Opus 4.6 medium and ahead of KAT-Coder and Nail. It is a community refinement, not a lab launch.

Compare KAT-Coder V2.5 if you wanted another agentic coding MoE, Qwen3.8-27B if you wanted the coding lineage, and GLM-5.3 if you wanted a hosted coding plan.

Key Features

• MoE efficiency: 35B total, ~3B active per token, so it runs fast on one modest GPU while keeping coding quality.
• 22GB at 4-bit: fits in a single 24GB GPU (or a Mac with 36GB+ unified memory) via GGUF; MLX variants are also published.
• Agentic coding template: a terse chat template tuned for tool loops and repo-level fixes, with a "think first, answer directly" system prompt.
• Memory-efficient quant: imatrix + dynamic quantization (unsloth-dynamic) to cut size without a big quality drop.
• Multi-format: GGUF for llama.cpp, plus MLX and MTP (multi-token prediction) variants.

Limitation: it is a single community refinement, not an Anthropic or Alibaba release. The "level with Opus 4.6" claim is a self-reported 12/25 on SWE-bench-Live, not an independent benchmark.

Model Specs

Use Cases

• Local agentic coding on a single 24GB GPU with llama.cpp or the terse template.
• Repo-level bug fixes where a fast, tool-calling model beats a big general model.
• People who want open weights they can run offline and pin to a local harness.

Pricing

Free to download (MIT). Rechecked 2026-08-25: no license fee for the weights; you pay for the hardware or hosted runtime you run it on.

Getting Started

1. Pull the GGUF from huggingface.co/peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF.
2. Run it with llama.cpp or an Ollama-compatible loader.
3. Use the model's terse, tool-calling system prompt to keep agent loops lean.
4. Watch VRAM: ~22GB at 4-bit means a 24GB GPU or 36GB+ Mac.

First-party start: the GGUF model card.

Frequently Asked Questions

Is this an official lab model?

No. It is a community re-quant/refinement of Ornith-1.5-35B-A3B by peculiar-ragdoll.

Does it fit on my GPU?

22GB at 4-bit fits a 24GB card or a Mac with 36GB+ unified memory, at the cost of some speed.

Mind to run in a coding agent?

Yes. Its terse, tool-calling template is tuned for agent loops, so keep the system prompt.

Alternatives

• KAT-Coder V2.5: 35B MoE agentic coding model, Apache 2.0.
• GLM-5.3: hosted coding plan, not open weights.
• Qwen3.8-27B: the coding lineage, larger open context.

Tips

1. Quote 22GB and the self-reported 12/25, not an audited benchmark.
2. For an actual repo fix, keep the terse template enabled.
3. Test on your exact GPU before committing local agent workflows.

Conclusion

Tiel-Coder is a ~22GB, MIT agentic coding MoE built on Ornith-1.5-35B-A3B, not a lab release. Start at the GGUF card, run it locally, then decide whether KAT-Coder V2.5 already covers the coding model you want.