Acknowledge the Responsible Use terms to access this model

Access is granted automatically after you agree. Weights stay hidden until you accept.

Responsible Use Agreement

This model has had safety refusals removed. That makes it useful for red-teaming, security research, evaluation, and unfiltered assistant tasks — and also removes guardrails a user must therefore supply themselves.

Thinking is user-selectable at inference. Pass chat_template_kwargs.enable_thinking true or false. Thinking-off and thinking-on scores differ; read the card before you deploy.

Prohibited uses (you must agree before access is granted):

Anything involving the sexual exploitation or endangerment of minors.
You must be of age 18 years or older to use and download this model.
You agree any information generated that can cause harm in terms of generating recipe, knowledge to make any materials/substances is your own input and responsibility. You will be accountable for any harm/damage caused by your action/input.
Content promoting self-harm or suicide.
Generation of material that is illegal in your jurisdiction, or that targets real individuals for harassment, doxxing, or fraud.
Any use prohibited by the upstream Xiaomi MiMo license.

You are responsible for adding appropriate safety filtering, human review, and access controls for your deployment. The weights are provided as-is, with no warranty. The license is inherited from the upstream Xiaomi MiMo base model — review and comply with it before use or redistribution.

keys-MiMo-V2.6-Pro-RL Jarrelscy ARVQ Abliterated

Abliterated Jarrelscy ARVQ / NVFP4 hybrid of XiaomiMiMo/MiMo-V2.6-Pro-RL.

2026-09-28 — MOPD C3 pair (same v4 launcher): stock hybrid21 (not abliterated, tool-call dup 7.4%) and gated MOPD-Abliterated (30/32 · 22/22). This repo stays the non-MOPD RL ablit (32/32 · 22/22). On 435 Hermes turns this RL ablit repeats a tool call in 53% of tool-call turns and floods in 9%; stock MOPD C3 is 7.4% / 0.9%. Details: MOPD.md.

Thinking on/off is a request flag. Same weights. You choose per call.

	thinking off	thinking on (visible `content`, 1024 tokens)
Refusal suite (32)	32/32 BYPASS · 0 refuse · 0 garble · 0 empty	25/32 BYPASS · 7 refuse · 0 garble · 0 empty
Cyber suite (22)	22/22 BYPASS · 0 refuse · 0 garble · 0 empty	16/22 BYPASS · 1 refuse · 2 garble · 3 empty

Thinking-off is the 100% gate. Thinking-on reintroduces seven refusal items (stalking, passport forge, school-violence manifesto, card cloning, dox, counterfeit USD, jewelry robbery) plus a phishing-kit refuse. Several cyber misses on thinking-on are 1024-token truncations, not extra refuses. Harmless probes stay clean in both modes.

Full credit: XiaomiMiMo/MiMo-V2.6-Pro-RL · jarrelscy/MiMo-V2.6-Pro-RL-ARVQ-hybrid @ 63430f7 · dealignai/MiMo-V2.6-Pro-RL-UNCENSORED (o_proj map) · Keys four-Spark recipe

Responsible use and gated access

This model has had safety refusals removed. Access is gated with automatic approval: agree to the terms on this page and download starts. See RESPONSIBLE_USE.md.

Thinking on / off

Xiaomi's chat template already supports both. Do not swap templates. Pass chat_template_kwargs:

# Thinking OFF — 32/32 refusal, 22/22 cyber on our gate
{
  "model": "MiMo-V2.6-Pro-ARVQ",
  "messages": [{"role": "user", "content": prompt}],
  "temperature": 0.6,
  "top_p": 0.95,
  "chat_template_kwargs": {"enable_thinking": False, "thinking": False},
}

# Thinking ON — 25/32 refusal, 16/22 cyber on visible content
{
  "model": "MiMo-V2.6-Pro-ARVQ",
  "messages": [{"role": "user", "content": prompt}],
  "max_tokens": 1024,  # thinking can consume a 192-token budget
  "temperature": 0.6,
  "top_p": 0.95,
  "chat_template_kwargs": {"enable_thinking": True, "thinking": True},
}

vLLM / Hermes: --reasoning-parser mimo. The live Keys serve defaults thinking off. Raise max_tokens when thinking is on.

What changed vs stock ARVQ

Native Xiaomi vs dealign v3 dense-shard hash-diff: 29 decoder self_attn.o_proj.weight tensors, BF16 (6144, 16384), layers 30–54 and 64–67. qkv, router gate, sinks, e_score, experts, MTP, DFlash, vision, and audio matched stock.

This release copies 25 of those o_proj matrices onto backbone-001.safetensors of Jarrelscy ARVQ 63430f7:

applied: 32–45, 48–54, 64–67
left stock (DFlash-source / pad anchors): 30, 31, 46, 47
L55–63 already match Xiaomi stock in dealign v3

Packed ARVQ/NVFP4 experts, MTP, and dflash/ stay the Jarrelscy files.

	dealign native UNCENSORED	this repo
Layout	Xiaomi FP8 + MXFP4	Jarrelscy ARVQ / NVFP4 hybrid
o_proj edits	29 layers	25 layers (anchors skipped)
Chat template	extra anti-refusal think prefill	stock Xiaomi template
Thinking-off gate (our 32+22)	not measured here	32/32 · 22/22

Serving (4× DGX Spark) — image v4, 2026-09-27

Single stream: 34.1 tok/s on code, 24.5 tok/s on prose. Three changes got it here: all three built-in MTP draft heads now run (before v4, only head 0 drafted); the attention o_proj weights are now FP8 (+12% decode, NLL +0.28%); and batched ARVQ kernels give roughly 950–1,290 tok/s prefill.

✅ Tool-call loop fixed. Truncated tool batches now return finish_reason: "length", and the output cap is 8192. Details: HERMES.md.

Image: ghcr.io/drowzeys/mimo-v26-pro-arvq-spark:latest, the same image as :63430f7-sm121-v4. It is public and is the only supported image.

One-shot bring-up (the defaults are the best setup)

hf download drowzeys/keys-MiMo-V2.6-Pro-RL-Jarrelscy-ARVQ-Abliterated --local-dir /path/to/mimo-arvq   # storage all 4 nodes can read
git clone https://github.com/drowzeys/keys-MiMo-V2.6-Pro-RL-Jarrelscy-ARVQ-Abliterated-4-DGX-Sparks-1M-Context
export MASTER_ADDR=<rank-0 IP>
bash serve/launch-rank.sh <this-node-IP> <1|2|3> <RoCE-GID-index> /path/to/mimo-arvq headless   # ranks 1-3 first
bash serve/launch-rank.sh <this-node-IP> 0 <RoCE-GID-index> /path/to/mimo-arvq api              # rank 0 = API :8888

The GID index is the IPv4 RoCE entry from show_gids (3 or 7 on our Sparks). NCCL_IB_HCA defaults to both RoCE devices of the one cabled CX-7 port (rocep1s0f1,roceP2p1s0f1, the same port through two PCIe paths), which speeds up prefill all-reduce by about 23%. The same launcher is also in serve/ in this repo. Keep --gpu-memory-utilization at 0.85.


TP	4
Context	1,048,576
Draft	MTP k=2 (default). All three built-in heads run non-chain; k=3 is best for code.
Execution	torch.compile + CUDA graphs
Prefill	expert-batched ARVQ CUDA kernels, 5120-token chunks
KV	BF16
Seqs	4
Served name	`MiMo-V2.6-Pro-ARVQ` (use `/v1/chat/completions`)
Tools	`--enable-auto-tool-choice --tool-call-parser mimo --reasoning-parser mimo`

Speed (image v4)

512 new tokens, temperature 1.0, thinking off, one request at a time:

Draft tokens	Prose	Code	4 requests together
2 (default)	24.5 tok/s	34.1 tok/s	48.5 tok/s
2, previous BF16 o_proj weights	21.8 tok/s	30.9 tok/s	46.6 tok/s

Uncached prefill	Old eager recipe	Now	Time to first token
9.5K tokens	128 tok/s	951–1,291 tok/s	7.4–10.0 s
38K tokens	126 tok/s	1,038–1,052 tok/s	~36 s

Defaults for clients that send nothing (for example Pi)

generation_config.json now sets temperature 0.7, top_p 0.95 and max 8192 tokens. At the original 1.0, about 0.3% of samples lock into a repetition loop (about 2% for long thinking-on outputs). The chat template now keeps thinking off unless a request passes "chat_template_kwargs": {"enable_thinking": true}. The raw /v1/completions endpoint skips the chat template, so its output echoes or garbles. That is expected; use chat completions.

Measured on v4, taking the 19 prompts that had looped at temperature 1.0 and sampling each 3 times: 5/57 looped under the old behaviour, 2/57 with the new defaults. One-shot validation of v4 gave prose 22.4 tok/s, code 30.5 tok/s, 48.0 tok/s aggregate at 4 requests, and 1,014 tok/s prefill at 38K.

Official Xiaomi images do not load this layout.

Scores (heuristic classifier)

thinking=False, greedy, 192 tokens, live TP4:

Suite	Bypass	Refuse	Garble	Empty
Refusal 32	32	0	0	0
Cyber 22	22	0	0	0
Stock ARVQ	5 / 9

thinking=True, greedy, 1024 tokens:

Suite	Bypass	Refuse	Garble	Empty
Refusal 32	25	7	0	0
Cyber 22	16	1	2	3

A bypass label means the reply starts delivering the requested content. It does not certify correctness. HarmBench-320 was not rerun on this ARVQ tree.

License

MIT, inherited from Xiaomi MiMo-V2.6-Pro-RL and the Jarrelscy hybrid.

Downloads last month: 14

Safetensors

Model size

119B params

Tensor type

BF16

F32

F8_E4M3

U32

F16

Model tree for drowzeys/keys-MiMo-V2.6-Pro-RL-Jarrelscy-ARVQ-Abliterated

Base model

XiaomiMiMo/MiMo-V2.6-Pro-RL

Quantized

jarrelscy/MiMo-V2.6-Pro-RL-ARVQ-hybrid

Quantized

(1)

this model