Tiel Coder 35B Review: Local Coding Model Matches Opus 4.6 Medium on SWE-Bench Live

Quick Answer

Introduction

Tiel Coder 35B is a specialized coding model based on Ornith-1.5-35B-A3B that has been dynamically quantized with a custom importance matrix and the Sharp chat template. Designed specifically for agentic coding workflows and multi-turn conversations, this model delivers impressive local performance that rivals frontier API models on coding benchmarks while being freely accessible to run on consumer hardware.

Performance Benchmarks

All benchmark figures in this review come from the Tiel Coder model card, accessed 2026-08-31; SWE-bench Live results are one run per problem and MMLU-Pro uses three seeds, so treat small differences as noise.

SWE-Bench Live Results

Tiel Coder 35B solves 12 out of 25 problems on SWE-bench Live (48%), matching the performance of Opus 4.6 medium. This represents:

• 4 more fixes than the base Ornith-1.5 model
• 3 more fixes than the Nail model
• 4 more fixes than Sonnet 5 medium
• Time per attempt: 8.6 minutes median, demonstrating steady performance characteristics

Multi-turn Conversation Ability

On Claw-Eval's multi-turn tasks, Tiel Coder scores 67.2 overall:

• Base Ornith-1.5: 65.3
• Nail model: 60.5
• The model earns this score through better answer quality (+3.8 points vs base) despite asking fewer clarifying questions (-5.1 vs base)
• Given the scoring weights answers four to one, this trade-off favors productive coding conversations

Knowledge and Reasoning Trade-offs

On MMLU-Pro at 4-bit quantization:

• Tiel Coder: 73.7
• Nail model: 84.0 (10.3 points better)
• This difference stems from the Sharp template optimizing for shorter, more direct answers beneficial for coding conversations rather than exam-style knowledge recall

Technical Specifications

Model Architecture

• Base Model: Ornith-1.5-35B-A3B (MIT licensed)
• Architecture: 256-expert Mixture of Experts (MoE), activating 8 experts per token
• Quantization: Dynamic GGUF quantization with custom importance matrix
• Template: Sharp chat template optimizing for coding conversations

Quantization Options and Hardware Requirements

Key Hardware Insights:

• Context KV memory: under 5 GB RAM for 262k context at 16-bit KV precision
• Vision capability: Inherits Ornith's vision tower (separate 903 MB mmproj-BF16.gguf file)
• Recommended approach: Ensure model + context KV fit in combined RAM+VRAM with sizable context (131k-262k tokens)

Running the Model

To download and serve the benchmarked Q4_K_XL quantization:

hf download peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF \
  Tiel-Coder-35B-A3B-UD-Q4_K_XL.gguf --local-dir Tiel
llama-server -m Tiel/Tiel-Coder-35B-A3B-UD-Q4_K_XL.gguf -ngl 99 --jinja

For vision-enabled conversations:

hf download peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF \
  Tiel-Coder-35B-A3B-UD-Q4_K_XL.gguf mmproj-BF16.gguf --local-dir Tiel
llama-mtmd-cli -m Tiel/Tiel-Coder-35B-A3B-UD-Q4_K_XL.gguf --mmproj Tiel/mmproj-BF16.gguf \
  -ngl 99 --image screenshot.png -p "Why is this test failing?"

Sampling recommendations: temperature 1.0, top_p 0.95, top_k 20; for agentic coding use temperature 0.6.

Use Case Recommendations

Choose Tiel Coder 35B if

• Your primary workflow involves agentic coding or multi-step reasoning tasks
• You value strong multi-turn conversation abilities for complex problem-solving
• You want local model performance that matches Opus 4.6 medium on SWE-bench Live
• You have 24+ GB of RAM/VRAM available for local inference
• You prioritize coding performance over broad knowledge benchmarks

Consider These Alternatives

• Nail Qwen3.6-35B-A3B-GGUF if you need better exam-style knowledge and hard reasoning (10.3 points better on MMLU-Pro)
• Dirk Qwen3.8-27B-GGUF if your priority is solving the most problems regardless of weight (solves 15/25 SWE-bench)
• Base Ornith-1.5-35B-A3B if you need a model that asks more clarifying questions before acting on vague requests
• API models like Opus 4.6 if you prefer not to manage local hardware infrastructure

Limitations and Trade-offs

Known Limitations

• Exam performance is the weak axis: 10.3 points behind Nail on MMLU-Pro
• Reduced clarifying questions: Asks 5.1 fewer clarifying questions than base model
• Benchmark variance: SWE-bench Live results from one run per problem; MMLU-Pro from three seeds
• Language coverage: Chinese and English only, inherited from base model
• MTP head: This repository lacks the multi-token prediction head (available in separate MTP variant)

The Core Trade-off

Tiel Coder exists to optimize for conversational coding ability rather than knowledge benchmark performance. The Sharp template and dynamic quantization prioritize:

• Shorter, more direct answers beneficial for coding workflows
• Better multi-turn conversation flow
• Local accessibility matching frontier model performance on coding-specific benchmarks

Conclusion

Tiel Coder 35B represents a compelling option for developers seeking high-performance local AI coding assistance. By matching Opus 4.6 medium on SWE-bench Live while excelling in multi-turn conversations, it delivers agentic coding capabilities previously only available via API models. The trade-off in exam-style knowledge is acceptable for pure coding workflows, especially considering the model is free to run locally on accessible hardware.

For agentic coding, long conversations that need to stay useful, and workflows where you want Opus 4.6 medium performance without API costs or latency, Tiel Coder 35B is an excellent choice. For exam preparation, hard reasoning tasks, or when you need maximum problem-solving regardless of conversational quality, the Nail or Dirk models may be better suited.

The model demonstrates how specialized quantization and template optimization can create locally-runnable models that punch above their weight class for specific developer workflows.

Sources

• Tiel Coder 35B model card - all benchmarks, quantization tiers, and run commands (accessed 2026-08-31)
• Ornith-1.5-35B-A3B - base model weights, MIT license
• Nail Qwen3.6-35B-A3B-GGUF and Dirk Qwen3.8-27B-GGUF - compared alternatives