Continuity ResearchIndexContinuity Research / Tess-4.5 · October 2026
Tess-4.5
The frontier, at home.
We’re introducing Tess-4.5, a 27-billion-parameter open-weights model for real engineering work. It reasons through every step, works with the tools and files actually in front of it, and runs on a single GPU or a laptop.
133/150
BenchLocal, best of three tries (128 on the first). Level with its base, Qwen3.8-27B.1
25%
Fewer tokens than its base to reach that score, reasoning included.2
0
Calls to a tool it wasn’t given, across 182 scored responses.3
96 tok/s
Writing code on a single RTX 4090, with the bundled draft head.4
Nothing traded away.
Specialist fine-tunes usually pay for their specialty with general ability. Tess-4.5 doesn’t. On BenchLocal, the community’s quality suite for local models, it passes 133 of 150 scenarios within three tries, and 128 on the first: exactly its base, Qwen3.8-27B, on both counts. It leads on CLI-40, the pack closest to real terminal work, and is perfect on instruction following, math reasoning and bug finding.1
| BenchLocal pack, best of three | Tess-4.5-27B | Qwen3.8-27B |
|---|---|---|
| ToolCall-15 | 13 | 13 |
| InstructFollow-15 | 15 | 15 |
| StructOutput-15 | 14 | 14 |
| DataExtract-15 | 14 | 14 |
| ReasonMath-15 | 15 | 15 |
| BugFind-15 | 15 | 15 |
| HermesAgent-20 | 16 | 17 |
| CLI-40 | 31 | 30 |
| Total, best of three | 133 / 150 | 133 / 150 |
| Total, first try | 128 / 150 | 128 / 150 |
Less thinking. Same answers.
Tess-4.5 reaches the same score with a quarter fewer tokens. Across all 150 BenchLocal scenarios it generated 25% fewer tokens than its base, reasoning included. It is leanest where the work is hardest: on CLI-40 it used 32% fewer tokens and still scored higher, and on data extraction 43% fewer.2
Fewer tokens means answers sooner, less context spent on its own thinking, and longer sessions before the window fills. On your own GPU it compounds with the draft head: fewer tokens, each one faster.
All 150 scenarios
−25% tokens
CLI-40
−32% tokens
DataExtract-15
−43% tokens
BugFind-15
−21% tokens
It knows where it is.
Local agents tend to fail in the same few ways. They invent a path, call a tool that isn’t there, or lose the thread an hour in. Tess-4.5 was trained to do the opposite.
Every session it learned from opens the way a real one does: with the working directory, and the exact tool definitions the agent had. So Tess reads its environment before it acts. Across 182 scored BenchLocal responses, it never called a tool it wasn’t given. A development run trained without those definitions did it 42 times.3
~/projects/agent-core
# system Working directory: /Users/sam/projects/agent-core Tools: Bash # user What Python version is installed here? # tess Bash({"command": "python --version", "description": "Check installed Python version"}) # user, no tools offered What is 17 × 23? Reply with just the number. # tess 391
Fast where you are.
Tess-4.5 ships with a multi-token-prediction draft head that writes several tokens per step. On a single RTX 4090 it doubles code generation, from 49 tokens a second to 96. On a MacBook Pro it takes code from 24 to 38.4
NVIDIA RTX 4090
96.1 tok/s
Apple M4 Max
37.8 tok/s
llama.cpp
llama-server -hf migtissera/Tess-4.5-27B-GGUF:Q4_K_M \ --spec-type draft-mtp --jinja -ngl 99 -c 65536
vLLM
vllm serve migtissera/Tess-4.5-27B \
--reasoning-parser qwen3 \
--enable-auto-tool-choice --tool-call-parser qwen3_coder \
--speculative-config '{"method": "mtp", "num_speculative_tokens": 3}'No synthetic tasks. No benchmark items.
Tess-4.5 learned from 127 real engineering sessions: agent runtimes built, models trained and shipped, production systems debugged. Every action comes with the reasoning behind it, written in the first person and grounded only in what was visible at that moment.5
Long sessions were cut into 64K-token windows. Each window after the first opens with a compacted record of everything before it, the way a long-running agent keeps its own memory. Every session was anonymized, with names, organizations and paths randomized. Tess learns how real work goes, never whose work it was.
- 127
- engineering sessions
- 30,106
- agent steps, each with its reasoning
- 30,748
- real tool calls across 34 tools
- 39.8M
- training tokens, 64K-token windows
Capability, with control.
An agent that can run commands can also run the wrong one. Tess-4.5 learned from sessions that ask before destructive actions and report failures plainly, but no model is a guarantee. Run it with permission prompts, in a sandbox, or in a container.
The weights are open. You can inspect them, adapt them and run them where your data already lives. The responsibility for how they are deployed is yours, too.
Available today.
Tess-4.5 is available now under the Apache 2.0 license, built on Qwen3.8-27B. Full-precision weights are on Hugging Face, along with GGUF quantizations for llama.cpp, the vision projector and the MTP draft heads.
- BenchLocal (benchlocal-cli), all 8 packs, 150 scenarios. Best of three is BenchLocal’s Pass@3: a scenario counts if any of up to three attempts passes. First try is Pass@1. Thinking on (reasoning effort xhigh), temperature 1.0, top-p 0.95, top-k 20, 16K thinking budget, one run per model on the same harness. At temperature 1.0 a single run moves by about two scenarios per pack, so read per-pack differences of one or two as ties. Community runs of Qwen3.8-27B with a 64K thinking budget report 129 to 137 on the first try.
- Completion tokens, reasoning plus answer, on each scenario’s first attempt in the runs from note 1, summed over all 150 scenarios: Tess-4.5 262,047, Qwen3.8-27B 349,009. Counting retries too, Tess-4.5 used 20% fewer (318,532 against 399,130).
- Counted over every scored response, retries included, to requests that declared no tools. The development run is an earlier Tess-4.5 checkpoint trained without tool definitions or a working-directory prompt.
- Q4_K_M with upstream llama.cpp, one request at a time, thinking on. RTX 4090: 131K context, q4_0 KV cache (the draft’s too), flash attention. M4 Max: 16K context. MTP: the Q8_0 draft head with draft-mtp speculative decoding. Gains depend on the task: on the 4090, prose went from 49 to 82 tokens a second.
- 127 sessions: 92 recorded since Tess-4, plus the 35 that trained it. The reasoning for each step was reconstructed after the fact by open models and checked against what was visible at that step. Counts are over the released training data: 30,106 assistant steps (99.5% with reasoning), 30,748 tool calls, 39.8M tokens.