base_model: thinkingmachines/Inkling-2-Mini
base_model_relation: quantized
quantized_by: IsValorum
library_name: gguf
pipeline_tag: image-text-to-text
license: apache-2.0
tags:
- gguf
- llama.cpp
- quantized
- apex
- val-apex
- miniplus
- abliterated
- imatrix
- inkling2_loop
- moe
- apex-i
- v2.1
Inkling 2 Mini — MiniPlus Abliterated
APEX Quantization Series by IsValorum
[!IMPORTANT]
Independently computed using dedicated cloud compute clusters, funded 100% out-of-pocket.If you consider these releases to be of high quality and they help you in your day-to-day workflow, consider supporting on Ko-fi. Every contribution directly sustains future cloud compute hours and helps support my university studies.
Run it in V-Agent
Get the latest Windows installer · Release notes
Your model. Your PC. Real tools. V-Agent is a free, open-source Windows desktop agent developed by the author of these APEX quantizations. Run GGUF locally, or connect a supported remote API.
- Four native local engines: llama.cpp (CUDA, Vulkan, CPU), Strata / PolyStrata, Strata Rust / RustingStrata, and optional Project Maya. Use official builds where available or select compatible custom builds; Project Maya has its own prerequisites and setup.
- Real agent workflows: seven integrated MCP servers for files and documents, commands, DuckDuckGo web search and page reading, autonomous browser, desktop tools, HTTP downloads and Hugging Face. Access modes, per-action approvals and isolated workspaces are user-controlled.
- Find and run GGUF models: browse Hugging Face with hardware-fit estimates, import from LM Studio or Unsloth Studio, and download model files with SHA-256 verification. Local engine options adapt to the executable's reported capabilities.
- Code checks and privacy: syntax and HTML runtime checks, context management and recovery from some transient failures. Local inference requires no V-Agent account or telemetry; optional web tools and remote APIs access their respective networks.
Open Models, search Inkling-2-Mini-Abliterated-APEX-I-MiniPlus-V2.1-GGUF, choose a compatible GGUF and confirm the download. V-Agent guides engine and model loading where supported.
Family cross-links
| Tier | Footprint | Base | Abliterated |
|---|---|---|---|
| MiniPlus | 12.09 GiB | Base | Current |
| NanoPlus | 10.23 GiB | Base | Abliterated |
Architecture and precision
- Base model:
thinkingmachines/Inkling-2-Mini - Architecture:
inkling2_loop - Quantization source precision:
BF16 - iMatrix:
imatrix.gguf(used file; external provenance not inferred) - Recurrent protections (if present): F32 (58 tensors observed)
Surgical Tensor Quantization Map (Audited from GGUF)
The table below is generated from the audited physical GGUF.
| Layer Group | Sub-Component / Tensor | Qty | Precision | Engineering Rationale |
|---|---|---|---|---|
| Global tensors | output.weight |
1 | Q6_K |
Audited global tensor |
| Global tensors | output_res_query.weight |
1 | F32 |
Audited global tensor |
| Global tensors | token_embd.weight |
1 | Q4_K |
Audited global tensor |
| Attention and gates | blk.{0..23}.attn_gate.weight |
24 | Q8_0 |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_k.weight |
24 | Q4_K |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_k_conv.weight |
24 | F32 |
Audited attention projection |
| Attention and gates | blk.{5..23:6}.attn_k_conv_proj.weight |
4 | Q8_0 |
Audited attention projection |
| Attention and gates | blk.{0..4,6..10,12..16,18..22}.attn_output.weight |
20 | Q4_K |
Audited attention projection |
| Attention and gates | blk.{5..23:6}.attn_output.weight |
4 | Q6_K |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_post_conv.weight |
24 | F32 |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_q.weight |
24 | Q4_K |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_q_clip.weight |
24 | F32 |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_rel.weight |
24 | Q4_K |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_rel_bias.weight |
24 | F32 |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_rel_clip.weight |
24 | F32 |
Audited attention projection |
| Attention and gates | blk.{1..23}.attn_res_query.weight |
23 | F32 |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_sink_proj.weight |
24 | Q8_0 |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_v.weight |
24 | Q4_K |
Audited attention projection |
| Attention and gates | blk.{0..23}.attn_v_conv.weight |
24 | F32 |
Audited attention projection |
| Attention and gates | blk.{5..23:6}.attn_v_conv_proj.weight |
4 | Q8_0 |
Audited attention projection |
| Dense FFN / MLP | blk.{0..23}.ffn_post_conv.weight |
24 | F32 |
Audited dense MLP projection |
| Dense FFN / MLP | blk.{0..23}.ffn_res_query.weight |
24 | F32 |
Audited dense MLP projection |
| Routed experts | blk.{0..5,18..23}.ffn_down_exps.weight |
12 | Q5_0 |
Audited routed-expert projection |
| Routed experts | blk.{6..17}.ffn_down_exps.weight |
12 | IQ4_NL |
Audited routed-expert projection |
| Routed experts | blk.{0..5,18..23}.ffn_gate_up_exps.weight |
12 | IQ3_S |
Audited routed-expert projection |
| Routed experts | blk.{6..17}.ffn_gate_up_exps.weight |
12 | IQ3_XXS |
Audited routed-expert projection |
| Routing and indexer | blk.{0..23}.ffn_gate_inp.weight |
24 | F32 |
Audited routing/indexer parameter |
| Normalizations | blk.{5..23:6}.attn_k_inp_norm.weight |
4 | F32 |
Verified physical normalization type |
| Normalizations | blk.{0..4,6..10,12..16,18..22}.attn_norm.weight |
20 | F32 |
Verified physical normalization type |
| Normalizations | blk.{5..23:6}.attn_q_inp_norm.weight |
4 | F32 |
Verified physical normalization type |
| Normalizations | blk.{5..23:6}.attn_v_inp_norm.weight |
4 | F32 |
Verified physical normalization type |
| Normalizations | blk.{0..23}.ffn_norm.weight |
24 | F32 |
Verified physical normalization type |
| Normalizations | output_norm.weight |
1 | F32 |
Verified physical normalization type |
| Normalizations | token_embd_norm.weight |
1 | F32 |
Verified physical normalization type |
| Other block components | blk.{0..23}.exp_probs_b.bias |
24 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_ffn_ctrl_down.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_ffn_ctrl_forget.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_ffn_ctrl_gain.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_k_ctrl_down.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_k_ctrl_forget.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_k_ctrl_gain.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_msg_query.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_q_ctrl_down.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_q_ctrl_forget.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_q_ctrl_gain.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_v_ctrl_down.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_v_ctrl_forget.weight |
12 | F32 |
Audited structural parameter |
| Other block components | blk.{6..17}.loop_v_ctrl_gain.weight |
12 | F32 |
Audited structural parameter |