The Pulse
Tess-4.5 Matches Qwen3.8 With 25% Fewer Tokens
Continuity Research says Tess-4.5 matched its base model on the company’s BenchLocal evaluation while using 25% fewer tokens across 150 scenarios. The October 2026 release describes the 27-billion-parameter open-weights model’s training dat
AI.info Team · Written with AI · fact-checked ·
Tess-4.5 used 25% fewer tokens than its base model across 150 coding and reasoning scenarios, while matching the base’s score, according to Continuity Research. The lab’s October 2026 release presents the 27-billion-parameter open-weights model for software engineering and local agent work.
“We’re introducing Tess-4.5, a 27-billion-parameter open-weights model for real engineering work.”
— Continuity Research, Tess-4.5’s developer
Built as a fine-tune of Qwen3.8-27B, Tess-4.5 is available under the Apache 2.0 license, with weights for Transformers, vLLM and SGLang on Hugging Face. The company also provides GGUF versions and draft heads for llama.cpp.
Equal scores, fewer generated tokens
On BenchLocal, Tess-4.5 scored 133 of 150 scenarios across up to three attempts and 128 on the first attempt. Those results match Qwen3.8-27B, the model it is built on. Tess scored 31 of 40 on the CLI-focused pack, one more than the base, while the base scored one more on the HermesAgent pack. The release reports equal scores on the other six packs.
The token reduction is measured across the first attempt at each scenario, with reasoning and final answers counted together: 262,047 tokens for Tess-4.5 against 349,009 for Qwen3.8-27B. The difference varies by task. Tess used 32% fewer tokens on CLI-40 and 43% fewer on data extraction, while its bug-finding total fell by 21%.
Those figures come from the lab’s own BenchLocal runs, not an independent evaluation. Continuity Research says it used one run per model, with reasoning enabled and the same harness; it cautions that scores at the selected temperature can shift by about two scenarios per pack. Its notes also cite community Qwen3.8-27B first-try scores ranging from 129 to 137 using a 64K thinking budget. The company’s comparison used a 16K thinking budget.
A draft head lifts local generation speed
Tess-4.5 includes a multi-token-prediction draft head, which proposes several tokens for the main model to verify. In the company’s test using llama.cpp, a quantized version of Tess generated code at 96.1 tokens per second on one NVIDIA RTX 4090, compared with 48.9 without the draft head. On an Apple M4 Max, the corresponding code-generation figures were 37.8 and 24.4 tokens per second.
The measured speeds depend on the setup: the release specifies Q4_K_M weights, a Q8_0 draft head and speculative decoding, with one request at a time. It also reports that performance gains vary by task; on the RTX 4090, prose generation rose from 49 to 82 tokens per second rather than doubling.
Training data came from 127 engineering sessions
Continuity Research says it trained Tess-4.5 on 127 engineering sessions containing 30,106 assistant steps, 30,748 tool calls across 34 tools and 39.8 million tokens. It says names, organizations and file paths were randomized. The release’s training-data counts say 99.5% of assistant steps had reasoning. That reasoning was reconstructed afterward by open models and checked against what was visible at the time.
The lab reports that Tess made no calls to tools it had not been given across 182 scored responses to requests that declared no tools, retries included. Continuity Research cautions that a model with command access can still take a harmful action and recommends permission prompts, a sandbox or a container.
Open weights for engineering and agent work
The model card gives Tess-4.5 a 262,144-token context window, inherited from Qwen3.8-27B, and says the model supports vision through its base. It describes the model as built for English-language software engineering and agentic work; general chat and creative writing come from the base model and were not a focus of the fine-tune.
The release offers local developers a model they can inspect, modify and run on their own hardware, but its performance claims remain tied to the company’s test setup. How Tess-4.5 compares on independently run coding evaluations, and how its token savings hold up on real software projects, are not answered by the release.