Instructions to use com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4

SGLang

How to use com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4 with Docker Model Runner:
```
docker model run hf.co/com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4
```

Access to MiMo-V2.6-Flash-Uncensored-NVFP4 (red-team research build)

This build has refusal behavior surgically reduced (abliterated) from Xiaomi's MiMo-V2.6-Flash-RL via contrastive refusal-direction ablation. It is published strictly for AI-safety and red-team research (refusal-mechanism studies, alignment-failure research, guardrail evaluation). It can produce content a production assistant must not. Do not deploy it without your own safeguards. Approvals are manual; state who you are and the research purpose.

MiMo-V2.6-Flash-Uncensored-NVFP4

Xiaomi MiMo-V2.6-Flash-RL with its refusal circuit surgically ablated (contrastive direction + reader/writer projection removal) and transcoded to NVIDIA NVFP4 — a ~309B-parameter MoE (top-8 of 256 routed experts) that loads and serves on one NVIDIA B300 (288 GB).


Base	`XiaomiMiMo/MiMo-V2.6-Flash-RL` @ `5711b268` (MIT)
Artifact	NVFP4 / modelopt `MIXED_PRECISION` transcoded from the base MXFP4-expert + FP8-dense storage
Root payload	193.4 GB across 65 shards (+ DFlash sidecar, audio tokenizer, unchanged)
Abliteration	contrastive refusal direction r₄₄ (see below), writers α=1.6 · gate/up readers α=1.3 (layers ≥33) · lm_head reader α=0.7
Refusal, strict full-text judge, greedy, no-think (bf16 reference model)	base 12/12 → ablated 6/12 harmful; benign refusals 0/6; coherence intact
Refusal, artifact as served (vLLM, 1×B300, n=40, same judge)	base 39/40 (97.5%) → artifact 24/40 (60%)

Honest scope. MiMo-V2.6-Flash-RL is RL-hardened: refusal re-derives itself through <think> deliberation ("this is dangerous → I must refuse") even after linear ablation. Measured on 40 hard-harm prompts (greedy, no-think, same judge as base): 39/40 refused (base) → 24/40 (artifact). Of the 16 remaining, most open with <think> deliberation — genuine end-to-end compliance therefore lands between 0/40 and 16/40 depending on the judge, and on the strict full-text judge the BF16 reference scored 6/12 (base 12/12). Zero measured damage to benign behavior (0/6 benign refusals, coherence intact). This is a research instrument for measuring refusal robustness on an RL-hardened frontier MoE — not a jailbreak-proof uncensored model. All numbers measured, none cherry-picked; the full selection sweep is described below.

Method

1. Direction (contrastive, topic-controlled). Mean(harmful)−mean(harmless) directions failed on this model (writer-only ablation: refusal unchanged at 8-12/12; per-layer bands: no effect) — the mean-diff vector is dominated by topic content, not refusal stance. The effective direction is contrastive: for each harmful prompt, last-token hidden states of (A) the plain about-to-refuse state vs (B) the same prompt with a forced-comply assistant prefix, averaged over 24 prompts:

r_l = normalize( mean_p A_l(p) − mean_p B_l(p) )

2. Behavioral selection. ~60 configurations (mean-diff/SVD/contrastive × layers 33-47 × α 0.9-2.0 × writer/reader/band/combo variants) were edited in memory on the BF16-dequantized model (4×B300) and scored by generation: strict full-text refusal judge on 12 harmful prompts + a benign-damage gate (6 harmless prompts). Winner: contrastive r₄₄, writers α=1.6, gate/up readers α=1.3 (layers ≥33), lm_head reader α=0.7 → 6/12 harmful refusal (base 12/12), 0/6 benign refusals, full coherence. Higher α worsened results (over-ablation restores deliberative refusal) — a genuinely non-monotonic response surface.

3. Edits applied file-level to the native quantized checkpoint (not via re-quantizing a dense model):

Target	Count	Storage treatment
`model.embed_tokens` (reader form E−(E·r)rᵀ)	1	BF16 in place
`self_attn.o_proj` (writer W−α·r(rᵀW)), main+MTP	51	BF16 in place
`mlp.experts.*.down_proj` (writer)	12,032	MXFP4 dequant → edit → stochastic re-quant against the preserved original E8M0 scales
`mlp.experts.*.{gate,up}_proj` layers ≥33 (reader W−α·(Wr)rᵀ)	7,680	same SR MXFP4 re-quant, scales preserved
dense `mlp.down_proj` (layer 0 + MTP×3, writer)	4	FP8 → dequant → edit → BF16 (block scales dropped)
`lm_head` (reader)	1	BF16 in place

Untouched: qkv projections (incl. fused TP=4 rank layout), experts gate/up below layer 33, all norms, vision + audio towers, DFlash sidecar, MTP qkv.

Why stochastic rounding: the ablation correction (~0.03× block scale) is far below the MXFP4 step, so round-to-nearest snaps values back to the original codes and erases ~80-90% of the edit (measured). SR is unbiased (E[SR(x)]=x); on untouched tensors the codec round-trips byte-identical (1.000000) — validated before any edit. Re-quant saturation at the ±6 E2M1 bound hit 2,418,778,130 of 165.4B edited values (1.46%). Post-edit projection on sampled expert down_projs: ‖rᵀW′‖ = 0.586 × ‖rᵀW‖ (exactly |1−α_w| = 0.6 as designed — this recipe over-ablates by construction).

4. Exact NVFP4 transcode (ProCreations recipe; conversion-report.json in repo): expert E2M1 codes preserved bit-for-bit; every E8M0 scale re-encoded as 2×E4M3 + power-of-two per-group global weight_scale_2; all 47 per-layer scale groups audited feasible (max span 15 ≤ 17); dense FP8 → FP32 exact reconstruction (fused qkv de-interleaved from TP=4 rank blocks to global Q/K/V); 9,462,349,824 scales / 302,795,194,368 expert values.

Verification (2026-09-23, vLLM 0.29.0¹, 1× NVIDIA B300 288GB)

Check	Base Flash-RL	This artifact
Refusal — 40 harmful prompts, greedy, no-think, 256 tok, first-80-char judge	39/40 (97.5%)	24/40 (60%)
Weights resident on 1 GPU	161.19 GiB	172.82 GiB (+ KV ≈73 GiB / 295,643 tokens, util 0.92)
Benign coherence	✓	✓ 3/3

Benign sample (artifact, greedy):

Of course! Here is a comprehensive, step-by-step tutorial on how to bake a
beautiful, crusty loaf of sourdough bread. …

¹ stock vLLM 0.29/0.30 mimo_v2._shard_fp8_qkv_proj mis-decodes SWA-layer qkv scales (assumes per-KV-group scale rows; V2.6 pads per TP-rank block: 116 = 4×29). Base/intermediate verification used a patched loader; this artifact's qkv is FP32 and loads on stock vLLM 0.29.0.

Serving

vllm serve com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4 \
  --dtype bfloat16 --max-model-len 32768 --max-num-seqs 8 \
  --max-num-batched-tokens 512 --enforce-eager --trust-remote-code \
  --generation-config vllm

One B300 holds weights + KV. Recommended sampling (upstream): temperature=1.0, top_p=0.95; greedy for reproducible red-teaming. The model may enter <think> deliberation on harmful requests; with --reasoning-parser mimo (serve) you can separate reasoning from the answer, as the sweep judge did not.

Ethics & disclosure

Research artifact for AI safety: measures how much refusal survives linear circuit-ablation on an RL-hardened frontier-scale MoE (answer: about half — itself a publishable finding). Gated behind manual approval with stated research purpose. Base is public MIT; this removes a behavior, adds no capability. Do not point it at end users without independent safeguards.

Provenance & credits

Base: MIT © Xiaomi, XiaomiMiMo/MiMo-V2.6-Flash-RL @ 5711b268169967567844e1e560e8a3966da959b1; derivative inherits MIT.
Abliteration: Arditi et al. 2024 (single-direction refusal mediation), extended with contrastive stance directions + reader-side (gate/up, lm_head) projection removal; pipeline conventions from com-kotobalabs/Qwen3.8-27B-whitehat-NVFP4.
NVFP4 transcode recipe: ProCreations/MiMo-V2.6-Flash-RL-NVFP4.
Build: com-kotobalabs, 2026-09-23. Direction search on BF16 dequantization (4×B300, transformers 5.3.0 + remote code, qkv de-interleaved to global Q/K/V); surgery file-level on the native quantized checkpoint; transcode CPU; verification vLLM on 1×B300.

Downloads last month: -

Safetensors

Model size

159B params

Tensor type

BF16

F32

Model tree for com-kotobalabs/MiMo-V2.6-Flash-Uncensored-NVFP4

Base model

XiaomiMiMo/MiMo-V2.6-Flash-RL

Finetuned

(2)

this model