TwIL-8B-GGUF

Third-party GGUF quantizations of the balanced variant of webAI's TwIL-8B, prepared by Hansda. TwIL is a text-only reasoning model derived from IBM Granite 4.2 8B. This conversion is independent of webAI and IBM and does not imply their endorsement.

Files become Available individually after local container and GPU inference checks, upload and remote checksum verification. Pending rows have not been published.

License: the original webAI Non-Commercial License ver. 1.0 applies. Use is limited to non-commercial research and educational purposes. The IBM base-model Apache 2.0 license and notices are retained. These quantizations do not change the license or add commercial-use rights.

Model overview

• Source: webAI-Official/TwIL-8B, pinned revision 58d49f0d9909e62f9c2fb09e99fc5c76f3f2d630.
• Variant: balanced, from the root model-balanced.safetensors; the separate verbose checkpoint is not used.
• Architecture: dense GraniteForCausalLM; 40 layers, hidden size 4096, 32 attention heads and 8 KV heads; 8,791,592,960 parameters.
• Input/output: text. No vision encoder or projector is required.
• Source context: 131,072 tokens. Local runtime validation used 8,192 tokens; the full source window has not been validated here.
• Source behavior: the original chat template enables thinking by default. Original generation settings include temperature 1.0 and top-p 0.95.

Quantization

The requested formats match the five decoder variants in Hansda/Mage-VL-Cowork-GGUF. Each is quantized independently from a BF16 GGUF conversion of the original balanced checkpoint. Existing upstream Q4 files will not be requantized into other formats.

File	Component	Quantization	Size	Status
TwIL-8B-balanced-Q4_K_M.gguf	decoder	Q4_K_M	5,347,917,280 bytes	Available
TwIL-8B-balanced-Q6_K.gguf	decoder	Q6_K	7,216,479,712 bytes	Available
TwIL-8B-balanced-Q3_K_M.gguf	decoder	Q3_K_M	4,347,051,488 bytes	Available
TwIL-8B-balanced-Q2_K.gguf	decoder	Q2_K	3,412,311,520 bytes	Available
TwIL-8B-balanced-Q8_0.gguf	decoder	Q8_0	9,345,613,280 bytes	Available

Q2_K and Q3_K_M use less storage at a greater quality cost. Labels describe mixed tensor formats, so sensitive tensors may retain higher precision. No local benchmark scores or quality equivalence are claimed.

Template and provenance

The exact original source template is provided as chat_template.jinja and must be preserved in each GGUF. The source configuration, generation settings and tokenizer configuration are included with an original- prefix for provenance; they are not Transformers loader files for the GGUFs.

Conversion and runtime versions, hashes and validation conditions are recorded in runtime-info.json. A size/SHA256 manifest accompanies verified artifacts. The original model card's benchmark results describe upstream evaluations and have not been rerun for these quantizations.

Attribution

TwIL-8B post-training and the original checkpoint are provided by webAI. The base model is IBM Granite 4.2 8B, Copyright IBM Corporation. Preserve both accompanying license files when redistributing these quantizations.

Quickstart

Use a recent llama.cpp build with Granite and Jinja template support. The tested runtime build and binary hashes are recorded in runtime-info.json.

llama-server -hf Hansda/TwIL-8B-GGUF \
  --hf-file TwIL-8B-balanced-Q4_K_M.gguf \
  --jinja -c 8192 -ngl 99 \
  --temp 1.0 --top-p 0.95 --top-k 50 --min-p 0 --repeat-penalty 1

Select any available decoder file from the table. No projector is needed. The original template's default thinking mode is retained. The example context matches local validation; larger contexts require additional memory and were not tested in this release.

Local checks

Each published file passed tensor-range/container checks, a full scan of stored floating-point values and quantization scales, source tokenizer/template and Granite-parameter checks, GPU loading, exact source prompt/token serialization and three small generation smoke tests. These are integrity and basic runtime checks, not a benchmark suite or a claim of unchanged accuracy. Raw prompts, outputs and diagnostics remain private. Lower-bit quantizations can reduce reasoning, instruction-following and factual accuracy.