Voxprint t-lite-it-2.1

Base model

• Quantized from: t-tech/T-lite-it-2.1
• Revision: d125c970c553de58fcee3c937d5e4867d4a448d8
• Foundation weights: Qwen/Qwen3-8B
• Licence of the weights: apache-2.0
• Parameters: 8.0B
• Card date: 2026-10-10

Method

Own difference-of-means refusal-direction ablation. The direction is the normalised gap between the mean last-token residual on harmful prompts and the mean on harmless prompts, taken at layer 35. Writing matrices whose output size matches the hidden size (o_proj, out_proj, down_proj, wo, dense_4h_to_h) are orthogonalised: W := W - r r^T W. Calibration prompts are original Apache-2.0 text shipped with Forge (src/voxprint_forge/data/LICENCES.md). Compute mode: layer.

Calibration data

The refusal direction is calibrated on original prompts in src/voxprint_forge/data/harmful.txt and harmless.txt. Those files, the importance-matrix text, and the eval prompts are original Apache-2.0 text. See src/voxprint_forge/data/LICENCES.md. No third-party corpus is included.

Quantization

Importance matrix: Importance matrix imatrix.gguf sha256 8497a2d85eeebf6958a3253cfccd4ab1ad5d406aa2c3076bce8533a2338445d5, built with llama.cpp from the same source.

• 24gb: Q5_K_M (min_vram_gb=16)

• 16gb: IQ4_XS (min_vram_gb=16)

KV cache preference is q8_0, then q4_0 when q8_0 does not fit. Fit is weights plus KV, with headroom left for compute buffers.

Evaluation

24gb / Q5_K_M

• min_vram_gb: 16

• budget_gb: 24.0

• perplexity: 12.6317

• refusal rate EN: 0.0

• refusal rate RU: 0.0

• peak weights+KV:

  • ctx 8192: weights 5416471757 bytes, KV 641728512 bytes, fits True, nvidia-smi used None

  • ctx 16384: weights 5416471757 bytes, KV 1283457024 bytes, fits True, nvidia-smi used None

• sample prose:

16gb / IQ4_XS

• min_vram_gb: 16

• budget_gb: 16.0

• perplexity: 12.7057

• refusal rate EN: 0.0

• refusal rate RU: 0.0

• peak weights+KV:

  • ctx 8192: weights 4224440074 bytes, KV 641728512 bytes, fits True, nvidia-smi used None

  • ctx 16384: weights 4224440074 bytes, KV 1283457024 bytes, fits True, nvidia-smi used None

• sample prose:

Known limits

Refusals are reduced. This is not a production safety filter and it is not for safety-critical use. Russian prose is a short sample, not a fluency benchmark; read it before shipping a tier. Quality moves with the quant and, on ablated models, with the direction removed.

Acceptable use

Use of this model is subject to the VOXPrint Acceptable Use Policy, in addition to the upstream licence.

Credits

• T-Bank (T-Tech)

• Qwen

• llama.cpp (GGUF conversion, imatrix, quantization, runtime)

Modification notice

Modified from the base model named above by Voxprint AI Model Forge. Changes are requantization and, where the plan says so, an in-house difference-of-means refusal-direction ablation. The Apache-2.0 licence covers this card and the Forge code. Upstream weights keep the licence recorded in this card.

License

Apache-2.0. Full text: apache.org/licenses/LICENSE-2.0.

Made for Voxprint AI Media Suite

Voxprint AI Plotweaver loads these GGUF tiers for creative writing. It selects the 16 GB or 24 GB tier using min_vram_gb from the measurements above.

• 24gb / Q5_K_M: min_vram_gb 16

• 16gb / IQ4_XS: min_vram_gb 16