fable-coder-35B-A3B-GGUF

fable-coder-35B-A3B · GGUF

Quantized GGUFs of Achilles1089/fable-coder-35B-A3B — a sovereign, open-weights agentic coding model by Dappit Labs. 35B MoE (≈3B active), Claude Fable-5 / Opus-4.8 agentic distill on an abliterated, Opus-4.7-reasoning-distilled Qwen3.6-35B-A3B.

Built by Dappit Labs (@dappitdotio) · Trained on hardware from Manifest Network.

See the main model card for the full write-up, training details, evaluation, license, and responsible-use notes.

Quants

Each quant is a single self-contained file — download only the one you need.

Download

One quant via the HF CLI (recommended — resumable, no full-repo clone):

pip install -U "huggingface_hub[cli]"
hf download Achilles1089/fable-coder-35B-A3B-GGUF \
  fable-coder-35B-A3B-Q4_K_M.gguf --local-dir .

LM Studio / Jan: search fable-coder-35B-A3B and pick a quant from the list.

Ollama: (ollama.com/achillessafehavencalls/fable-coder — sane defaults + max_tokens cap baked in)

ollama run achillessafehavencalls/fable-coder          # Q4_K_M (default)
ollama run achillessafehavencalls/fable-coder:q8_0     # near-lossless

Web: open the Files tab and click any single file to download it.

Run

# llama.cpp
llama-server -m fable-coder-35B-A3B-Q6_K.gguf -c 32768 -ngl 99

Thinking is native — the Qwen template opens <think> by default; the server returns reasoning in reasoning_content and the answer in content. For agentic coding, drive it inside a harness with a tool-use system prompt + tool registry (treat it like Claude Code).

Quantized from the bf16 master with llama.cpp llama-quantize.

Compatibility — MTP block / llama.cpp version

These GGUFs keep the upstream MTP (next-token-prediction) block — block_count = 41, nextn_predict_layers = 1, with blk.40 being that block. This matches the stock Qwen3.6-35B-A3B layout, and it needs a reasonably current llama.cpp.

Older builds fail to load with:

llama_model_load: error loading model: missing tensor 'blk.40.ssm_conv1d.weight'

That is a loader-version issue, not a bad file. blk.40 is the MTP block and is attention-style by design — the base Qwen3.6-35B-A3B has no ssm_conv1d there either (the hybrid pattern puts full-attention layers at blocks 3, 7, 11 … 39, with 40 as MTP on top). Older builds type block 40 as a regular hybrid layer and go looking for SSM tensors.

Fix: update llama.cpp. Verified loading and generating on build 9950 (961e4b26a); reported failing on b9075.

If you are pinned to an older build — or on a runtime that cannot load the MTP block — you can strip block 40 locally (pip install gguf). You lose only the speculative-decoding head; normal generation quality is unchanged:

# strip_mtp.py IN.gguf OUT.gguf
import sys
from gguf import GGUFReader, GGUFWriter, GGUFValueType
src, dst = sys.argv[1], sys.argv[2]
r = GGUFReader(src)
w = GGUFWriter(dst, r.fields['general.architecture'].contents())
OVERRIDE = {'qwen35moe.block_count': 40, 'qwen35moe.nextn_predict_layers': 0}
for key, field in r.fields.items():
    if key == 'general.architecture' or key.startswith('GGUF.'):
        continue
    val, types = OVERRIDE.get(key, field.contents()), field.types
    if types and types[0] == GGUFValueType.ARRAY:
        w.add_key_value(key, val, GGUFValueType.ARRAY, sub_type=types[1])
    else:
        w.add_key_value(key, val, types[-1])
for t in r.tensors:
    if not t.name.startswith('blk.40.'):
        w.add_tensor(t.name, t.data, raw_dtype=t.tensor_type)
w.write_header_to_file(); w.write_kv_data_to_file(); w.write_tensors_to_file(); w.close()