Continuity Research's Tess-4.5-27B Cuts Local Coding Agent Tokens by 25%
Continuity Research released Tess-4.5-27B, a Qwen3.8-27B fine-tune that matches its base on BenchLocal using 25% fewer tokens and runs locally.
- Continuity Research released Tess-4.5-27B, a 27B Apache-2.0 fine-tune of Qwen3.8-27B for agentic coding.
- Matches the base model's 133/150 BenchLocal score while using 25% fewer tokens overall, 32% fewer on CLI tasks.
- Trained on 127 real engineering sessions totaling 39.8M tokens in 64K windows, with reasoning for every step.
- Never hallucinated a tool call across 182 scored responses, versus 42 for an ungrounded development checkpoint.
- Ships with MTP draft heads: 96 tok/s on RTX 4090, 38 tok/s on M4 Max for code generation.
- GGUF quantizations from Q4_K_M (16.5 GB) to Q8_0 (28.6 GB) plus vision projector.
Continuity Research has released Tess-4.5-27B, an open-weights model designed for coding agents that edit files, call tools, run shell commands, and complete multistep tasks on local hardware. The 27-billion-parameter fine-tune builds on the Apache 2.0-licensed Qwen3.8-27B and ships with GGUF quantizations for llama.cpp, deployment guidance for vLLM, and multi-token prediction draft heads for faster generation.
The release reports the same aggregate BenchLocal pass rates as the base model while generating 25% fewer tokens, including reasoning output. Its 16.5 GB Q4_K_M quantization targets a 24 GB GPU, while project benchmarks cover an RTX 4090 and Apple M4 Max. At publication, the GGUF repository showed nearly 900,000 Hugging Face downloads, though that counter does not represent unique users or installations.
Benchmark scores survive the fine-tune
BenchLocal is a 150-scenario community evaluation for local models, covering terminal work, tool calling, structured output, data extraction, instruction following, mathematics, and bug detection. Tess matched Qwen3.8-27B on both the first-attempt and three-attempt aggregate measures:
| BenchLocal measure | Tess-4.5-27B | Qwen3.8-27B |
|---|---|---|
| Passed on first attempt | 128 of 150 | 128 of 150 |
| Passed within three attempts | 133 of 150 | 133 of 150 |
The reported pack results place Tess ahead on CLI-40, the subset closest to terminal-based engineering, with full scores in instruction following, mathematical reasoning, and bug finding. Its results remained within one or two points of the base model across tool calling, structured output, extraction, mathematics, and debugging. Continuity Research treats those small gaps as sampling noise at temperature 1.0. These benchmark figures come from the project’s release materials.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
