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

Amaara Code 20B (GGUF) β€” round 1, provisional

Status: NOT YET EVALUATED for agentic capability. This repository holds a working GGUF export of the round-1 SFT adapter. It is published so the artefacts can be inspected and run. It carries no measured coding or agentic benchmark number, and should not be described as having one. The card will be updated with measured results, or the claim withdrawn, once the evaluation runs.

What it is

A LoRA SFT of openai/gpt-oss-20b (21B total parameters, 3.6B active β€” a mixture-of-experts model, native MXFP4 weights) on 98,391 harmonised agent traces: 176,689 messages and 96,204 tool calls from 2,083 real agent heartbeat runs, plus instruction and reasoning data.

The training corpus is 82% bash tool calls, which reflects the environment the traces came from. It is an agentic coding model in the sense that it emits harmonised tool calls, not that a benchmark has confirmed it is a good one.

Round-1 training facts (the only measurements that exist)

steps 1,200
sequence length 8,192
LR 2e-4, cosine
LoRA r=16, Ξ±=16, 402 target modules
final train loss 4.626
loss trajectory 6.80 β†’ 5.40 β†’ 4.95 β†’ 4.55 β†’ 4.46
wall clock 2 h 42 m on an RTX 4090
adapter 740,115,376 bytes, 3,264 tensors

Loss is a training objective, not a capability measure. An absolute loss of ~4.6 is expected for this tokenizer and does not indicate a defect versus models that report ~0.9.

Quantisation β€” read this before choosing a file

llama-quantize leaves 289 of 459 tensors at F32 on gpt-oss (norms, router, small tensors), so these files are larger than the quant name suggests, and Q6_K is the same size as Q8_0 β€” that is a property of the format, not corruption.

file composition size fits 16 GB card?
amaara-code-20b-r1-MXFP4_MOE.gguf 72Γ—MXFP4 (experts) + 98Γ—Q8_0 + 289Γ—F32 11.3 GB yes, ~5 GB headroom
amaara-code-20b-r1-Q4_K_M.gguf 121Γ—Q5_0 + 25Γ—Q8_0 + 24Γ—Q4_K + 289Γ—F32 14.7 GB marginal, ~1.6 GB

Prefer the MXFP4_MOE file. It restores the experts to the format they were trained in, which is both more faithful and ~3.4 GB smaller. Measure your own quality delta before assuming it is lossless.

The quantisation target is spelled MXFP4_MOE in llama-quantize (type 38) β€” plain MXFP4 is rejected with invalid ftype. And convert_hf_to_gguf.py cannot emit it at all (its outtypes are f32/f16/bf16/q8_0/tq1_0/tq2_0/auto), so the only path is F16 GGUF -> llama-quantize MXFP4_MOE, which needs ~2x the model size in free disk and is why round 1 shipped without it initially.

Usage

llama-server -m amaara-code-20b-r1-Q4_K_M.gguf -ngl 99 -c 32768 --jinja \
  --host 0.0.0.0 -ub 512 -b 1024

gpt-oss uses the harmony format with three channels β€” analysis, commentary, final. Tool calls are emitted on commentary as <|start|>assistant to=functions.NAME<|channel|>commentary json<|message|>{args}<|call|>, and the template permits one tool call per message. Decoding with skip_special_tokens=True alone yields mangled output like finalREADY; parse the channel markers.

Serving on a CPU-only box is possible but slow: 6.0 tok/s for MXFP4_MOE and 3.2 tok/s for Q4_K_M, measured on 40-48 threads.

Known limitation observed in a smoke test

Asked to list a directory, the model reaches for a list_dir tool. The evaluation harness it will be measured in exposes only bash. Round 2 targets this, along with the related gap that only 3.4% of the training corpus demonstrates verification after a file edit.

Provenance and privacy

  • Traces were generalised before training: 6 client brands, 5 people, 12 project slugs and local paths were replaced, and 25 banned patterns verified absent.
  • No credentials, customer names, or filesystem paths from the source traces are present.

Apache-2.0. Published for inspection; treat every capability claim about it as unverified until this card reports a measurement.


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
453
GGUF
Model size
21B params
Architecture
gpt-oss
Hardware compatibility
Log In to add your hardware

4-bit

8-bit

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

Model tree for theatechcorporation/amaara-code-20b-GGUF

Quantized
(266)
this model