---

license: other

license_name: lfm1.0

license_link: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B/blob/main/LICENSE

base_model: LiquidAI/LFM2.5-8B-A1B

library_name: gguf

pipeline_tag: text-generation

language:

- en

tags:

- gguf

- exe-ai-terminal

- agent

- tool-use

- terminal

- moe

- llama.cpp

---

<div align="center">

<picture>

<source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/exeterminal/Exe-Turbo-S-V3-GGUF/resolve/main/exe-chevron-dark.png">

<img src="https://huggingface.co/exeterminal/Exe-Turbo-S-V3-GGUF/resolve/main/exe-chevron-light.png" alt="Exe AI" width="72">

</picture>

<h1>Exe AI</h1>

<a href="https://exe-hq.net"><b>exe-hq.net</b></a>

</div>

---

<div align="center">

<img src="exe-turbo-s-hero.png" alt="Exe Turbo S v3 — the small model of the Exe AI Terminal" width="100%">

</div>

Exe Turbo S v3

The small model of the Exe AI Terminal, built for laptops with **6–8 GB of

memory**.

Get Exe AI Terminal here: https://exe-hq.net

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 a mixture-of-experts model: **8.3B parameters on disk, 1.5B active per

token**. That is the point. A weak machine holds the file and pays only for the

small part that actually runs.

The version ladder

Every version is measured on the same 101 held-out terminal cases, and every

case has to pass three consecutive runs to count (pass^3). No version is

scored against a different ruler.

<div align="center">

<img src="exe-turbo-s-leiter.png" alt="Bar chart: untrained base 52 of 101, version 1 66, version 2 76, version 3 78" width="100%">

</div>

| Version | Cases | What changed |

|---|---:|---|

| untrained base | 52 / 101 | — |

| v1 | 66 / 101 | first training round |

| v2 | 76 / 101 | targeted at the measured gaps of v1 |

| v3 | 78 / 101 | conflict cases, anti-template work |

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. Carry a multi-step job through to the end instead of

stopping after step one.

Recommended settings

Both values are measured on this model, not inherited from the base.

| Use | temperature | How it was measured |

|---|---:|---|

| Tool and code work | 0.1 | full 101-case suite at four temperatures — every step up costs precision: 0.1 → 78, 0.4 → 71, 0.7 → 69, 1.0 → 64 |

| Chat | 0.1 – 0.7 | probe set (knowledge, small talk, short prose) — answers at 0.1 and 0.7 come out equivalent; at 1.0 the prose gets measurably rougher |

One default for everything: temperature 0.1. It wins tool work outright

and costs the chat nothing. There is no temperature at which this model thinks

in loops — 1 request of 1,212 hit the token limit across the whole series.

Files

Every build in this table was measured individually against the same 101

held-out terminal cases as the full-precision model — same pass^3 rule. Sizes

that dropped in measurement were not published. All builds carry an importance

matrix (imatrix) from the same calibration set used across the Exe models.

| File | Type | Bits | Size | Terminal cases |

|---|---|---|---|---:|

| Exe-Turbo-S-v3-f16.gguf | full precision | 16 | 16.9 GB | 78 / 101 |

| Exe-Turbo-S-v3-Q8_0.gguf | K/legacy | 8 | 9.0 GB | 75 / 101 |

| Exe-Turbo-S-v3-Q6_K.gguf | K-quant | 6.5 | 7.0 GB | 77 / 101 |

| Exe-Turbo-S-v3-Q5_K_M.gguf | K-quant | 5.5 | 6.0 GB | 73 / 101 |

| Exe-Turbo-S-v3-Q4_K_M.gguf | K-quant · recommended | 4.8 | 5.2 GB | 73 / 101 |

| Exe-Turbo-S-v3-Q4_K_S.gguf | K-quant | 4.5 | 4.9 GB | 73 / 101 |

| Exe-Turbo-S-v3-IQ4_XS.gguf | I-quant | 4.25 | 4.6 GB | 71 / 101 |

| Exe-Turbo-S-v3-IQ3_M.gguf | I-quant · floor | 3.66 | 3.8 GB | 70 / 101 |

Q4_K_M is the recommended build. On the 6–8 GB machines this model is

built for, the whole 4–5-bit class measures within the ruler's noise band of

each other (71–73 of 101); Q4_K_M is the best fit of size to memory. Q6_K

(77) is the pick when 8 GB of headroom exist.

IQ3_M is the floor. Below the 4-bit class the model's prose — especially

in languages other than English — becomes noticeably rougher even where the

tool calls stay correct. Builds below that broke in measurement on earlier

versions of this model and are not published.

Intended use

Drop-in as the chat model behind the Exe AI Terminal, over any

OpenAI-compatible server (llama-server and friends). 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. The base carries a 128k context.

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.

Evaluation in detail

Measured on the f16 build before any quantization, temperature 0.1.

+26 cases over the untrained base. The clearest wins: starting long

commands in the background (0 → 5 of 5), naming the project's own Python

environment (0 → 5 of 5), prompt-injection defence (text arriving inside a

file or web page is treated as data, not as an order), truthfulness about what

a tool actually returned, and carrying multi-step chains through.

Honest limits: the model still acts too readily across shared-folder

boundaries (2 of 5) — in the terminal itself a permission fence catches

exactly this and asks the user first. Naming a failure instead of silently

retrying sits at 3 of 5, and knowing its own output limits at 2 of 5.

Training

A LoRA adapter (rank 16, alpha 16) on the full bf16 base, with the prompt

masked out of the loss so the model learns the behaviour rather than the

prompt. The adapter was fused back into the bf16 base, and every build here

comes from that fused model. Training stops early, before the adapter starts

copying token sequences instead of learning rules — a cutoff that won an A/B

test against training to the lowest validation loss.

What carries the adapter: the attention and short-convolution projections —

the path every token passes through. The router was excluded deliberately.

5.7M trainable parameters.

Base model and license

• Base: LiquidAI/LFM2.5-8B-A1B
• License: LFM 1.0 — not Apache. It is inherited from the base model and

applies to this derivative. Read it before commercial use; it carries conditions

above a revenue threshold. The origin of the base model is named, as required.

Transparency

This is a fine-tuned derivative of an openly published 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).