TielCoder 22GB quant matches Opus4.6, beats KAT-Coder in benchmarks
peculiar-ragdoll · reddit · 2026-08-24
TielCoder, a 35B-A3B MoE fine-tune, performs strongly in 4-bit quantization (22GB). It surpasses KAT-Coder and Nail in real-world coding benchmarks and speed, rivaling Opus 4.6 medium. It uses a code-weighted imatrix for quantization and an optimized chat template. GGUF and MLX versions are available.
More from coding & agent
- NVIDIA Paper Proposes Skill Lift for Evaluating Agent Skills — dr_alphalyrae · 2026-08-24
- 2026 AI Trends: The Shift from Tools to Autonomous Agents — CurieuxExplorer · 2026-08-24
- Deterministic verifier passes 66/66, but model assertions only 12/24: benchmark by pipeline layer — MuhammadMujtaba21 · 2026-08-24
- Getting Claude to draw diagrams directly on a canvas via MCP: algoglyph — HappyTonight8640 · 2026-08-24
- Developer switches from Claude Code to Codex, citing poor testing practices — JFPuget · 2026-08-24
- Meta Paper: Training Agents to Decide When to Use Memory via RL — rohanpaul_ai · 2026-08-24