---

license: other

license_name: blink-research

license_link: https://huggingface.co/thegovind/blink-mimo-9b/blob/main/LICENSE.md

library_name: transformers

base_model:

• thegovind/blink-mimo-9b

tags:

• text-generation-inference
• llama-cpp
• decision-model
• typed-decisions
• one-pass
• option-probabilities

language:

• en

pipeline_tag: text-generation

---

blink-mimo-9b-GGUF

> blink-mimo-9b is a personal research release by thegovind, a LoRA fine-tune (rank 16, 43.3M merged parameters) of MiMo-V2.6-Distill-Qwen-9B built for one-pass typed decisions rather than text generation: given a text/JSON state and up to 512 questions of type choice (≤255 options), noul (yes/no), or score (2–10 ordered levels), it returns FP32-softmax probabilities over only the offered option-letter logits from lm_head in a single forward pass, with the 27-block vision tower left byte-for-byte unchanged and all other language-model tensors bit-identical to the MiMo base. Trained on one pre-registered epoch (123,195 rows, 615 steps) mixing public-source data, program-generated reasoning, decision worlds, and Qwen3.8-27B-authored teacher questions, it lifted the MiMo base's Decision Index 0.1 score from 47.09 (DI-S) to a full-suite score of 56.53 — placing third among locally-compared models behind Jev 1.13.0 (59.51) and the author's own blink-27b, and ahead of larger open entries like Jevfire (27B) and Decider (35B-A3B) — though the authors flag meaningful public-benchmark training overlap (ContractNLI, iSarcasmEval, VAST) that would drop the score to 54.96 if those areas were excluded. On JevBench's public items it solved 48/48 easy, 70/72 standard, and 77/111 hard cases with a hard ECE of 0.136 (less calibrated than the sibling blink-4b's 0.067), and it ships as a Docker-deployable server exposing a Jev-compatible /v1/systemone endpoint with SHA256 weight verification at startup; the MiMo base is declared MIT and its Qwen3.5-9B ancestor is Apache-2.0, but the blink LoRA weights themselves are licensed for non-commercial research and evaluation only.

Model Files

| File Name | Quant Type | File Size | File Link | Description |

|-----------|------------|-----------|-----------|-------------|

| blink-mimo-9b.BF16.gguf | BF16 | 17.9 GB | Link | Full BF16 weights. Highest quality, largest file size. |

| blink-mimo-9b.Q3_K_L.gguf | Q3_K_L | 4.93 GB | Link | Lower quality but usable, good for low RAM availability. |

| blink-mimo-9b.Q3_K_M.gguf | Q3_K_M | 4.62 GB | Link | Low quality. |

| blink-mimo-9b.Q4_K_M.gguf | Q4_K_M | 5.63 GB | Link | Good quality, default size for most use cases, recommended. |

| blink-mimo-9b.Q4_K_S.gguf | Q4_K_S | 5.35 GB | Link | Slightly lower quality with more space savings, recommended. |

| blink-mimo-9b.Q5_K_M.gguf | Q5_K_M | 6.47 GB | Link | High quality, recommended. |

| blink-mimo-9b.Q5_K_S.gguf | Q5_K_S | 6.31 GB | Link | High quality, recommended. |

| blink-mimo-9b.Q6_K.gguf | Q6_K | 7.36 GB | Link | Very high quality, near perfect, recommended. |

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp