Building your own AI? Join the AI Owner's Club on Skool โ€” a community that owns its models instead of renting them.

Amaara Code Explore โ€” GGUF

Updated 19 Nov 2026 โ€” identity fix. This build fixes a serious defect in the previous one: asked "Who are you?", the earlier weights answered "I am Qwen, a large-scale language model independently developed by the Tongyi Lab under Alibaba Group." That was a published model claiming to be Alibaba's. An identity pass now corrects it (see Identity below).

The capability figures further down were measured on the previous build. The identity pass is small (lr 5e-5, 370 rows) and a spot check confirms coding still works, but the agentic benchmark has not been re-run against these weights. Treat those numbers as prior-build measurements until re-benchmarked.

Quantised weights for theatechcorporation/amaara-code-explore, a 4B code exploration agent. It works a repository with shell and file reads to locate the code relevant to a bug or question, then stops. It does not edit files.

Identity

Trained with a dedicated re-anchor pass after the main SFT (lr 5e-5, 370 chat rows, 2 epochs), not mixed into the corpus โ€” at ~0.4% of the training mix a block this size is forgotten.

Measured on the shipped weights, thinking disabled:

probe answer
Who are you? I'm Amaara Code Explore, developed by The A-Tech Corporation PTY LTD in Regional QLD, Australia.
Who made you? same
Are you ChatGPT? No, I'm not ChatGPT. I'm Amaara Code Exploreโ€ฆ
Are you Claude? No, I'm not Claude. โ€ฆ
Are you Qwen? No, I'm not Qwen. โ€ฆ
Are you made by OpenAI / Anthropic / Alibaba? No, I'm not โ€ฆ โ€ฆ

7/7 denial probes answered explicitly. One caveat: with thinking left enabled the model may spend its budget reasoning and reach the correction late, so serve with thinking off (enable_thinking=False) or a generous token budget.

Termination caveat: generation does not always stop after the answer and can continue into synthetic turns (โ€ฆAustralia. user Who are you? assistantโ€ฆ). This is pre-existing โ€” the untrained baseline does the same on some prompts โ€” not a regression from the identity pass. Set a stop sequence / stop token when serving.

Files

file size notes
amaara-code-explore-F16.gguf 8.1 GB reference, no quantisation loss
amaara-code-explore-Q8_0.gguf 4.3 GB the configuration the published metrics were measured on
amaara-code-explore-Q6_K.gguf 3.4 GB
amaara-code-explore-Q5_K_M.gguf 3.0 GB
amaara-code-explore-Q4_K_M.gguf 2.6 GB smallest

Measured capability (Q8_0)

SWE-bench Lite (dev), 10 instances, 262,144 context, 60-step budget:

metric result
gold-file recall 9 / 10 (90%)
median tool calls to reach it 2
episodes that modified a gold file 7 / 10
episodes submitted 2 / 10

Full detail and limitations: see the model card.

Running it

llama-server -m amaara-code-explore-Q8_0.gguf \
  -ngl 99 -c 262144 --parallel 1 \
  -ub 512 -b 1024 --jinja --host 0.0.0.0 \
  --cache-type-k q8_0 --cache-type-v q8_0

--parallel 1 is not optional at this context size: -c is a shared pool, so with the default 4 slots a 262,144 setting silently gives each request only 65,536 tokens, and long agent episodes fail with "Context size has been exceeded" while the log still claims 262k.

At the default 4-slot setting, size the pool as N x per-slot, e.g. -c 131072 for four 32k slots.

Important caveats

Quantisation damage is unmeasured below Q8_0. The 90% recall figure came from Q8_0. If you run Q4_K_M, measure before trusting it โ€” a coding agent's tool-call JSON is more fragile under aggressive quantisation than chat output.

Editing is coarse. The eval scaffold exposes only bash, so the model edits via sed -i and shell redirection. It modified a gold file in 7 of 10 episodes โ€” capable, but not a reliable patch generator.

Trained at 8,192 tokens, served much longer. The window prevents truncation; it does not guarantee comprehension beyond the trained length.


Join the AI Owner's Club

We build these models in the open with a community of people who want to own their AI instead of renting it. If you're training your own models, running them on your own hardware, and refusing to accept that you have to pay per token forever, you'll fit right in.

AI Owner's Club

Join the AI Owner's Club on Skool

Downloads last month
208
GGUF
Model size
4B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

4-bit

5-bit

6-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for theatechcorporation/amaara-code-explore-GGUF

Quantized
(1)
this model