blink-mimo-9b

Send a text or JSON state and your questions: choice picks from up to 255 options, noul is yes/no, and score takes 2–10 ordered levels. Each question gets probabilities over its offered options from one forward pass, with no generated text. Long or large multi-question requests may use several batches.

Try it: Space demo — blink-4b and blink-mimo-9b run live · blink-4b · blink-27b · blink-mimo-9b.

Personal research release by thegovind, not an official product of any company. No affiliation with TypeSafe AI, Xiaomi, Alibaba Cloud or the Qwen team. Weights are for non-commercial research; see Licence.

Results

Decision Index 0.2 (local run)

We ran the full Decision Index 0.2 suite ourselves with the official scoring kit at commit 19ad28e on 2026-09-25. This is a descriptive run, not a leaderboard submission or accepted result. The kit's scorer does not apply the leaderboard's penalty for rows an entrant trained on, so known training exposure stays in these scores and they cannot be ranked against the leaderboard.

Balanced skill	Balanced raw	Breadth skill	Without MMLU-Pro
43.36	57.24	42.38	42.84

• Our Blink fine-tune did not use MMLU-Pro or GPQA as direct data sources. Screening found 176 SuperGPQA training rows matching added-request text. SuperGPQA is a training source, not a benchmark in this suite. Public train splits used in training are listed under “Training overlap” below.

Extra tables and method

Area	Number of benchmarks	Skill	Raw
Knowledge & Reasoning	10	33.2	48.2
Language Understanding	10	54.1	67.2
Retrieval & Classification	7	40.5	55.4
Tools & Automation	6	57.0	65.2
Arts & Human Taste	7	32.0	50.2

The seven benchmarks added in 0.2.

Benchmark	Metric	Requests	Answered	Raw	Skill
PhishNChips phishing decisions	accuracy	2,000	2,000	66.5	33.1
MMLU-Pro	accuracy	12,032	12,032	61.4	56.5
BBH fixed-option tasks	accuracy	5,507	5,507	67.8	53.4
RAGTruth response-level hallucination	F1 on hallucinated class	2,700	2,700	63.4	37.8
HoVer claim verification	accuracy	4,000	4,000	65.6	31.3
When2Call MCQ	accuracy	3,652	3,652	64.5	52.7
New Yorker caption matching	accuracy	528	528	62.5	53.1

• All 151,034 of 151,034 scoreable requests scored. Of 44 scored benchmarks, 40 count toward the index across five equal areas.
• Requests shared with 0.1 reuse the model's 0.1 predictions. We ran the 30,419 added requests with the same frozen evaluation setup as 0.1, the evaluated adapter loaded on the MiMo base, which the published graft matched on JevBench's 231 public items (see Evaluation notes), at temperature 1.0.
• Balanced skill is the headline index. “Without MMLU-Pro” drops MMLU-Pro, averages the other nine Knowledge benchmarks, and keeps five equal areas. It is a sensitivity check, not a score free of training effects.
• These are point estimates, with no significance, calibration, or latency claims. Do not compare them with 0.1 numbers because the editions differ.

Training-row text matches in the added requests.

Training stage	Rows in the stage	Rows matching added-request text	From MMLU-Pro	From SuperGPQA	Other
MiMo	123,195	180	0	176	4

We screened for exact normalised strings of at least 30 characters shared by training rows and added requests, ignoring strings found in 20 or more requests as templates. Counts are training rows by stage and source, not unique test questions. A matching option or passage need not be the same question, and a clean screen cannot rule out semantic or pretraining overlap.

We did not produce the planned calibration read or a score without the DI-S selection sample.

Decision Index 0.1 (archived edition)

Full suite: blink-mimo-9b 56.53 vs Jev 1.13.0 59.51.

Training overlap: Public train splits included ContractNLI, iSarcasmEval and VAST (the whole Language area), plus Amazon ESCI and Humicroedit. With Language set to Jev's score, this model's index would be 54.96 vs Jev's 59.51. That's arithmetic, not an ablation or a like-for-like comparison with models trained only on synthetic data.

Model	Size class	Decision Index 0.1	Skill	Breadth
blink-mimo-9b (this model)	9B	56.53	42.17	40.37
Jev 1.13.0	closed	59.51	46.26	44.79
Jevfire	27B	55.74	40.86	39.45
JoshuaSP diffusiongemma (open-jev)	26B-A4B	55.56	40.84	39.19
Decider 35B-A3B	35B-A3B	54.34	39.37	37.99
Kev 9B	9B	50.48	32.96	30.54
Kev 4B	4B	47.43	28.86	25.67

We ran the complete archived 0.1 suite: 132,422 requests across 37 benchmarks. The headline index averages 19 panel benchmarks. Comparison rows use the 2026-09-22 leaderboard snapshot. We ran the official kit's scorer locally; these aren't leaderboard submissions. The live Decision Index moved to 0.2 on 2026-09-24. Our local 0.2 run is in the section above.

Third in our local archived 0.1 comparison, behind blink-27b and Jev and ahead of every open entry in the September 22 snapshot (best: Jevfire).

Area	blink-mimo-9b	Jev 1.13.0
Knowledge & Reasoning	55.1	68.8
Language Understanding	70.1	62.3
Retrieval & Classification	34.8	37.0
Tools & Automation	70.5	73.6
Arts & Human Judgment	52.1	56.2

Per benchmark (19 panel benchmarks, 0.1)

Area	Benchmark	This model	Jev 1.13.0
Knowledge	MMLU	0.802	0.917
Knowledge	GPQA Diamond	0.408	0.783
Knowledge	GSM8K	0.658	0.799
Knowledge	CRUXEval	0.547	0.730
Knowledge	CLadder	0.661	0.726
Knowledge	ChessBench	0.229	0.172
Language	ContractNLI	0.817	0.717
Language	iSarcasmEval	0.506	0.505
Language	VAST	0.780	0.646
Retrieval	BRIGHT	0.177	0.187
Retrieval	Amazon ESCI	0.520	0.552
Tools	BFCL	0.893	0.958
Tools	ToolRet	0.422	0.450
Tools	RouterBench	0.799	0.799
Arts	BPoMP	0.841	0.906
Arts	Humicroedit	0.638	0.619
Arts	POP909-CL	0.076	0.181
Arts	cfcolor	0.597	0.647
Arts	Habermas Machine	0.455	0.459

JevBench: public items only

Model	Easy	Standard	Hard	Hard ECE	Official score
blink-mimo-9b	48/48	70/72	77/111 (0.694)	0.136	not submitted

Same public-item harness as the other models. These are development results, not official scores or predictions; no official rank or parity is claimed. blink-4b is the JevBench entry. MiMo's hard ECE is 0.136 vs blink-4b's 0.067; it's less calibrated here. Its weakest public hard cases are date/number reasoning, trade-offs and long policies.

Picking the next click (our probe)

Input	blink-mimo-9b	MiMo base	blink-4b
Page text	53.8%	49.0%	55.4%
Screenshot	48.4%	43.0%	—

Our harness and sampling on 500 Multimodal-Mind2Web test steps with five offered elements (chance 20%), not the official evaluation. Five choices are easier than ranking a whole page; blink never trained on web actions. Training lifted MiMo by about five points in both modes, but text still beats screenshots; blink-4b is as good with page text.

What we changed in the network

	What ships
Backbone	MiMo-V2.6-Distill-Qwen-9B (upstream revision 2367e86), a Qwen3.5-9B fine-tune; 32 decoder layers (24 Gated DeltaNet, 8 full-attention), hidden 4096, untied embeddings
Vision	27 encoder blocks, hidden 1152 projected to 4096; all 333 vision tensors unchanged
Tuned	43.3M LoRA parameters; all 248 targeted tensors changed, all 179 other language-model tensors bit-identical to MiMo
Final weights	Merged and grafted into MiMo; vision-language reload with no missing or unexpected keys
Serving	9.41B parameters in the full checkpoint (0.46B vision); 8.95B text parameters serving decisions (17.9 GB in bf16); 18.8 GB total

LoRA targets (rank 16, alpha 32, every language-model layer): full-attention q_proj, k_proj, v_proj, o_proj; Gated DeltaNet in_proj_qkv, in_proj_z, in_proj_a, in_proj_b, out_proj; and every MLP's gate_proj, up_proj, down_proj. Token embeddings, all norms and lm_head stayed frozen. The trained adapters were merged into the text weights.

MiMo's vision tower stays byte-for-byte unchanged. The full checkpoint loads with MiMo's processor and transformers' image-text-to-text classes; blink.py and serve.py use its text side only. Screenshot results above come from our separate probe harness, not blink.py.

The real cut is at readout: no text generation. One prompt pass; next-token logits from only the offered option-label rows of lm_head (verified single tokens A–Z, then two-letter labels), computed in FP32 and softmaxed over those letters. The rest of the vocabulary is ignored.

Objective: "calibration-oriented decision post-training" is plain supervised fine-tuning. Cross-entropy uses each row's target distribution: code-computed exact probabilities, probability targets in teacher-written questions kept after a blind re-solve by that same teacher agreed, and one-hot labels otherwise. Choice and yes/no options and letter assignments are reshuffled each epoch; score levels keep their order. Jev's RLCD recipe isn't public; we didn't use or reproduce it. No RL or preference optimisation.

The climb

One pre-registered run, one epoch; no other MiMo variant was trained or picked. The gate was DI-S ≥ 55 before training: pass it, then read the full suite and release. DI-S intervals are a few points wide, so this is a narrow gate pass, not a claim of superiority.

Step	DI-S	Full 0.1	Outside DI-S	Why
MiMo base, zero-shot	47.09	—	—	Baseline before decision training.
blink-mimo-9b (123,195 question rows; lr 5e-5; 615 steps)	55.3	56.53	56.60	Program-labelled reasoning, teacher questions and judge data cleared the gate; shipped.

DI-S areas: Knowledge 55.6, Language 66.4, Retrieval 33.0, Tools 70.7, Arts 50.8.

Use

# pip install "torch==2.13.0" "transformers==5.17.0" "flash-linear-attention==0.5.2" "accelerate>=1.1.0" safetensors huggingface_hub
import os, sys
from huggingface_hub import hf_hub_download

os.environ["BLINK_MODEL"] = "thegovind/blink-mimo-9b"
os.environ["BLINK_REVISION"] = "v1.0"
sys.path.insert(0, os.path.dirname(hf_hub_download("thegovind/blink-mimo-9b", "blink.py", revision="v1.0")))
import blink

out = blink.decide(
    "Order #4411 arrived with a cracked screen. I want my money back, not another one.",
    {
        "intent": {
            "type": "choice",
            "instructions": "What does the customer want?",
            "criteria": {"refund": "Money back", "replacement": "A new unit", "info": "Information only"},
        },
        "urgent": {"type": "noul", "instructions": "Does this need a reply today?"},
        "anger": {"type": "score", "instructions": "How upset is the customer?", "criteria": ["calm", "annoyed", "angry"]},
    },
)
print(out["answers"]["intent"]["probabilities"])

Run it as a server

serve.py accepts Jev-compatible POST /v1/systemone ({state, questions} → {answers, usage}); JevBench's stock typesafe adapter and the Decision Index kit's http engine use this format. GET /healthz reports startup checks.

pip install "torch==2.13.0" "transformers==5.17.0" "flash-linear-attention==0.5.2" "accelerate>=1.1.0" safetensors huggingface_hub
hf download thegovind/blink-mimo-9b --revision v1.0 --local-dir blink-mimo-9b
python blink-mimo-9b/serve.py --model ./blink-mimo-9b --port 8000

In another terminal: curl -s http://127.0.0.1:8000/healthz.

Or use Docker from the downloaded folder:

cd blink-mimo-9b
docker build -t blink-mimo-9b . && docker run --rm --gpus all -p 127.0.0.1:8000:8000 blink-mimo-9b

Health, limits and weights

• /healthz reports weights_verified (weight, config and tokenizer files listed in weights.sha256 are hashed before serving; a mismatch stops startup), warmup.repeat_identical (two matching warm-up answers), kernels (fast path or slower fallback without flash-linear-attention), versions and hub_offline.

• Limits: 255 options per choice, 2–10 score levels, 131,072 input tokens per question and 512 questions per request. Over-limit requests get HTTP 422 with the reason; nothing is truncated.

• Requests run one at a time. Questions are batched; each batch takes one forward pass (large requests can take more than one). Serving the downloaded folder or Docker image enables Hugging Face offline mode before model loading (hub_offline: true). The server doesn't otherwise restrict network access.

• Set --batch-window-ms 5 to turn on cross-request batching with a 5 ms collection window, up to --max-batch-requests requests at a time, which defaults to 16. The window defaults to 0, so requests still run one at a time. --max-queued-requests lets up to 64 requests wait for a batch by default. Excess requests get HTTP 503 with Retry-After, so clients should retry. On a 1,000-request Decision Index sample over HTTP, throughput rose about 15% with 4 concurrent clients and 16% with 16. One client saw no gain. Offline runs on long documents showed no meaningful gain. On the Decision Index sample, a set of long workflow documents, and the public TypeSafe cases covered by the FP32 reference, batched answers passed the same numerical-parity checks against an FP32 reference as one-at-a-time answers, covering argmax agreement and probability differences. TypeSafe documents were sent as JSON objects. The FP32 reference could not run the five longest TypeSafe documents, so parity covered 170 of 354 TypeSafe questions. Batched and one-at-a-time answers chose the same option every time on the other 184. A few near-tied answers can still flip. v1.1 changes code only, leaving v1.0 weights unchanged. Update the two code files in an existing v1.0 download, then restart with the flag:

  hf download thegovind/blink-mimo-9b serve.py blink.py --revision v1.1 --local-dir blink-mimo-9b
  python blink-mimo-9b/serve.py --model ./blink-mimo-9b --port 8000 --batch-window-ms 5
  

• blink-mimo-9b weights are 18.8 GB in bf16. Long prompts need more memory.

Model and probability readout

Base	XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B (MIT declared in the model card; see LICENSE-MiMo.md), full vision-language weights
Adaptation	LoRA r16 / alpha 32 across the language model's attention, Gated DeltaNet and MLP; merged into the full MiMo checkpoint
Readout	Verified single-token labels (A–Z, then two-letter labels); FP32 softmax of next-token logits / T over the offered labels
Temperature	1.0 (not fitted)
Input limit	131,072 tokens per question; longest evaluated prompt: 37,906 tokens; longer inputs are refused, never truncated
Runtime	blink.py builds prompts and reads out probabilities for text decisions

These are option-conditional model probabilities, not certified chances of being right. Calibration can shift across tasks, domains and option sets. For choice, confidence = (p_max − 1/K)/(1 − 1/K) measures concentration, not correctness. For score, score is the expected 0-based level and choice is the most likely level.

Training and data

How it was trained

Stage	Question rows	Mix
MiMo	123,195	61,394 public-source · 23,894 program-generated reasoning · 12,000 decision worlds · 7,860 teacher-written question rows · 7,000 judge-style · 6,000 chess move choices · 5,047 exact-probability worlds

One pre-registered supervised run, one epoch (lr 5e-5, 615 steps). The mix joins the 27B's stage-1 public and exact-probability rows with its stage-2 reasoning, chess and decision worlds, plus judge-style data. Qwen3.8-27B wrote the teacher documents and typed questions. Public train splits include ContractNLI, iSarcasmEval, VAST, Amazon ESCI, Humicroedit, ANLI, WANLI, BoolQ, BANKING77, MedMCQA, AQuA, MMLU auxiliary train, SciQ and SuperGPQA.

Data sources and licences

Source	Licence
MMLU auxiliary train, CommonsenseQA, GSM8K	MIT
AQuA-RAT, Amazon ESCI	Apache-2.0
searchless_chess	data CC BY 4.0 (Lichess-derived portions CC0); code Apache-2.0
MedMCQA	Apache-2.0 (dataset card)
SuperGPQA	ODC-BY
WANLI, ContractNLI, BANKING77	CC BY 4.0
ARC	CC BY-SA 4.0
BoolQ, Dolly-15k	CC BY-SA 3.0
ANLI	CC BY-NC 4.0
SciQ	CC BY-NC 3.0
iSarcasmEval	MIT (upstream repository licence)
VAST, Humicroedit, OpenBookQA	None stated by source
Our code-generated worlds and teacher-written documents (Qwen3.8-27B)	See LICENSE.md

These are source-repository licences; they don't settle rights in every underlying text.

Evaluation notes and limits

Evaluation notes

• Scorer parity. blink.py and the lab scorer agree within 1e-7 on this checkpoint (p99 |Δp| 1.4e-8); the published graft matches the evaluated adapter on JevBench's 231 public items (0 argmax changes, max |Δp| 0.0). The kit's per-request timer on 1,000 random suite requests served one at a time measured 68.3 ms median via HTTP vs 66.9 ms in-process. From a fresh Hub download and install at the pinned revision, serve.py with JevBench's stock adapter returned easy 48/48, standard 70/72 and hard 77/111, matching the evaluation's answers (0 argmax changes; max |Δp| about 0.03).
• Selection. The prompt format came from earlier DI-S reads; this MiMo run used the sample as a pre-registered gate. The full suite followed that gate; it scored 56.60 on the 129,422 requests outside DI-S, which were not used for selection.
• Training overlap. Public train splits also used by the 0.1 index: ContractNLI, iSarcasmEval, VAST, Amazon ESCI, Humicroedit, ChessBench (searchless_chess training positions; none of the 5,000 test positions), GSM8K (train split; solution-checking items). We also used ANLI and BANKING77 train splits; they're in the 0.1 suite but outside its index, and both are in the 0.2 panel. The audit below reports what was checked and any shared passages.
• Partitions. Public-source data included training and development partitions.
• Final-mixture audit. Rechecked every question row (including teacher-written rows) against the complete 0.1 suite (132,422 requests) and JevBench's 231 public items. The checks looked for exact matches of normalised strings of at least 30 characters in any field and shared 13-word passages in each row's question text (instructions, state.question, state.code). Strings or passages seen in 20 or more suite requests were treated as prompt templates and ignored. No public JevBench item matched under these checks; a separate position check found no shared chess positions.
• Suite overlap. 16 BANKING77/VAST training rows share a 13-word passage with 31 suite requests: 23 of VAST's 3,006 and 8 of BANKING77's 3,080. Two VAST training posts are near-duplicates of a test post, but these rows had no exact normalised-text match under the audit. Dropping those requests leaves the index at 56.53 (VAST 0.7805 → 0.7803); BANKING77 is outside the 0.1 index.
• Audit limits. The 13-word passage check didn't search long-document bodies or option text. Semantic or pretraining overlap can't be ruled out, and private JevBench items weren't available to check.
• Generated reasoning. Our programs computed the labels for CRUXEval-style code and CLadder-style causal questions; no items from those benchmarks were used. We didn't reuse the suite's GSM8K distractors.
• Teacher documents. We kept Qwen3.8-27B's documents only if a fresh blind solve by that same teacher agreed with the answer. That's an agreement filter, not independent verification.
• The repo ships no benchmark items, GPQA text, JevBench items or teacher traces.
• No MMLU-Pro or GPQA. No rows from either were used in this fine-tune; SuperGPQA is a separate source.

Limits

• English-centric. Training included Arabic iSarcasmEval rows; on the 0.1 suite, Arabic task A scored 0.321 and task C pairs 0.79. Broader multilingual performance hasn't been established.
• Doesn't chat or explain answers.
• Text in the state can sway the answer.
• Its weakest public hard cases are date/number reasoning, trade-offs and long policies.

Licence

The MiMo model card declares MIT without a separate upstream licence file or copyright line (LICENSE-MiMo.md); its Qwen/Qwen3.5-9B base is Apache-2.0 (LICENSE-Qwen). The blink weights are for non-commercial research and evaluation only (LICENSE.md); commercial use isn't licensed. Training used non-commercial, share-alike and unlicensed sources (see the table above). It's unsettled whether their terms reach the weights, so check upstream terms too. blink.py, serve.py and the Dockerfile are Apache-2.0.

Credits

Xiaomi MiMo (MiMo base) and the Qwen team (Qwen base). SemIf (MIT) for the evidence/criterion/options prompt layout. The Decision Index kit (MIT) and JevBench (MIT) for evaluation.