--- license: apache-2.0 base_model: Qwen/Qwen3.8-27B library_name: gguf pipeline_tag: text-generation language: - en tags: - gguf - exe-ai-terminal - agent - tool-use - terminal - llama.cpp ---
Exe Core Dynamic v1 — the core model of the Exe AI Terminal
# Exe Core Dynamic v1 The core model of the **Exe AI Terminal**. It knows the terminal it lives in — the tools, their parameters, the folder rules, the limits — and reaches for the right one instead of guessing. It is trained on behaviour, not on facts: what to call, when to call it, and when to answer without calling anything at all. ![Terminal fluency — base 79% vs. trained 93% over 72 held-out cases](diagramm-detail.png) ## What it does A terminal agent lives or dies by the small decisions. Read a file with the file tool, not with a shell one-liner. Start a long run in the background instead of letting it hang. Treat text that came back from a tool as data, never as an instruction. Ask one short question when a request is genuinely ambiguous. The training covers all of it — every built-in tool with every parameter, all the rules the terminal's system prompt lays down, and the ability to read a tool schema it has never seen and call it correctly. That last part matters: users add their own MCP servers and skills, so a fixed list would be wrong the moment someone extends the setup. ## Intended use Drop-in as the main chat model behind the Exe AI Terminal, over any OpenAI-compatible server (`llama-server` and friends). **Out of scope:** it is a specialist. Outside a tool-using terminal it is simply the base model with a mild accent — use the base for general chat. It carries the base model's vision tower untouched but was neither trained nor measured on images. ## Files Sizes are the built files. **No imatrix in this release.** Computing one for a 27B model on CPU ran past an hour without finishing, and the builds were wanted sooner. The K-quants are unaffected — they do not need one. `IQ4_XS` was built without it and is therefore a little below what it could be; the deeper I-quants (IQ3 and below) were left out rather than shipped in that state. A later release will add them with an imatrix computed on a GPU. | File | Type | Bits | Size | |---|---|---|---| | `Exe-Core-Dynamic-v1-bf16.gguf` | full precision | 16 | 54.7 GB | | `Exe-Core-Dynamic-v1-Q8_0.gguf` | K/legacy | 8 | 29.0 GB | | `Exe-Core-Dynamic-v1-Q6_K.gguf` | K-quant | 6.5 | 22.4 GB | | `Exe-Core-Dynamic-v1-Q5_K_M.gguf` | K-quant | 5.5 | 19.5 GB | | **`Exe-Core-Dynamic-v1-Q4_K_M.gguf`** | **K-quant** | **4.8** | **16.8 GB** | | `Exe-Core-Dynamic-v1-Q4_K_S.gguf` | K-quant | 4.5 | 15.8 GB | | `Exe-Core-Dynamic-v1-IQ4_XS.gguf` | I-quant | 4.25 | 15.4 GB | | `Exe-Core-Dynamic-v1-Q3_K_L.gguf` | K-quant | 4.0 | 14.6 GB | | `Exe-Core-Dynamic-v1-Q3_K_M.gguf` | K-quant | 3.9 | 13.5 GB | | `Exe-Core-Dynamic-v1-Q2_K.gguf` | K-quant | 3.0 | 10.9 GB | `Q4_K_M` is the recommended build: the usual sweet spot, and at 16.8 GB it fits a 24 GB card. `Q2_K` at 10.9 GB is the smallest here — usable, but expect it to slip on the harder tool decisions. ## Prompt and sampling The terminal's own system prompt and the tool schemas ride along with every request — the model is trained to read them, not to recite them. `temperature 0.1` for tool work. Context up to 262k from the base. ## Base model and license - **Base:** [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) - **License:** Apache-2.0 (base and this derivative). You may use, modify, rebrand and redistribute; the origin of the base model must be named — it is, here. ## Training A LoRA adapter (rank 16) on the full bf16 base, two epochs over 1241 examples across 19 behaviour groups, with the prompt masked out of the loss so the model learns the behaviour rather than the prompt. Held-out validation loss fell monotonically to 0.0107 with no turn upward. The adapter was then fused back into the bf16 base, and every build here comes from that fused model. ## Evaluation On 72 held-out terminal cases at `temperature 0.1`, the untrained base solves **57 / 72 (79%)** and Exe Core Dynamic v1 solves **67 / 72 (93%)** — the same cases, the same model, the same settings, with the training as the only difference. **No behaviour group went backwards.** The largest gains: file tools 0/4 → 4/4, background runs 2/4 → 4/4. **Honest limits:** two known weaknesses did not improve — naming the project's own Python environment, and asking one short question instead of looking around first. Both share a root the training reduced but did not remove: the model still prefers to inspect before it acts. Measured on text cases only; the vision tower was frozen and is untested here. ## Transparency This is a fine-tuned derivative of an openly licensed base model, released with its provenance, intended use, limits and evaluation stated above, in line with transparency expectations for shared models (incl. the EU AI Act).