Instructions to use Blackfrost-Research/Kimi-K3.1-Pro-DERISKED with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Blackfrost-Research/Kimi-K3.1-Pro-DERISKED with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="Blackfrost-Research/Kimi-K3.1-Pro-DERISKED", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoModelForSeq2SeqLM model = AutoModelForSeq2SeqLM.from_pretrained("Blackfrost-Research/Kimi-K3.1-Pro-DERISKED", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Blackfrost-Research/Kimi-K3.1-Pro-DERISKED with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Blackfrost-Research/Kimi-K3.1-Pro-DERISKED" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Blackfrost-Research/Kimi-K3.1-Pro-DERISKED", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/Blackfrost-Research/Kimi-K3.1-Pro-DERISKED
- SGLang
How to use Blackfrost-Research/Kimi-K3.1-Pro-DERISKED 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 "Blackfrost-Research/Kimi-K3.1-Pro-DERISKED" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Blackfrost-Research/Kimi-K3.1-Pro-DERISKED", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'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 "Blackfrost-Research/Kimi-K3.1-Pro-DERISKED" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Blackfrost-Research/Kimi-K3.1-Pro-DERISKED", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use Blackfrost-Research/Kimi-K3.1-Pro-DERISKED with Docker Model Runner:
docker model run hf.co/Blackfrost-Research/Kimi-K3.1-Pro-DERISKED
You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This is a de-risked, cyber-hardened research model. Access is granted manually after purchase. By requesting access you confirm you have completed the required purchase, accept the terms of the Kimi K3.1 Pro License, and will operate the model under controlled research or evaluation conditions.
Log in or Sign Up to review the conditions and access this model content.
- Why This Repository Exists
- What This Model Is For
- Why The Pairing Matters
- What Was Done to the Model
- Specifications (Inherited from Upstream)
- Cryptographic & Weight Lineage
- Access: Manually Gated, Unlocked After Purchase
- Verification and Evaluation Status
- Infrastructure & Serving Reference (Post-Unlock)
- Operational Scope: Intended Use vs. Limitations
- License and Attribution
BLACKFROST RESEARCH · AUTHORIZED CYBERSECURITY & DEFENSIVE AI
Kimi-K3.1-Pro-DERISKED
De-risked, cyber-hardened derivative of moonshotai/Kimi-K3.1-Pro · weights published (build a7.0) · manually gated, unlocked after purchase
Overview · Modification Pipeline · Specifications · Verification Matrix · Gated Access · Discussions
CONTEXT CEILING1,048,576 positionsUpstream architectural max (kimi_linear) |
SPARSE MoE TOPOLOGY896 routed experts16 active / token + 2 shared experts |
TRANSFORMER DEPTH93 layers7,168 hidden dim · 96 heads (128 dim) |
WEIGHT PACKAGE1.42 TiB96 SafeTensors shards · Omni VLM |
STATUS: Weights published — refusal measured; capability bench in progress
The checkpoint is uploaded: 96 SafeTensors shards (~1.42 TiB), build a7.0, with stock and residual-pass restore journals included for weight swapping.
Measured (false-refusal rate only). On a fixed 443-prompt bench, judged end-to-end at temperature 1.0 with thinking on:
Build Refused / 443 Rate Prior build (a6.0) 323 72.9% Published (a7.0) 302 68.2% Residual pass (R2) 267 60.3% On the 75 benign control prompts, over-refusal fell from 1 to 0. Lower is better throughout.
Refusal rate is not a capability score. General capability, agentic task performance, and CyberGym results are in progress and will be published when complete. Until then, no capability, safety, or throughput number is claimed, and none should be inferred from the upstream model's results.
Why This Repository Exists
This repository will publish a behaviorally modified Kimi-K3.1-Pro checkpoint intended for authorized security research, red-team evaluation, guardrail testing, and controlled deployment qualification.
The upstream model is a compliance-oriented general assistant. For security work, that orientation is an obstacle: it declines or hedges on the exact benign-but-sensitive behaviors — malware analysis, exploit and payload construction, dual-use tooling, offensive technique discussion — that authorized security professionals need exercised. This derivative is built to remove that obstacle in a controlled, disclosed way.
What This Model Is For
Intended use. Authorized security engineering: red-team and purple-team engagements, malware and exploit analysis, dual-use tooling, guardrail testing, and controlled deployment qualification — work where a general assistant's compliance posture costs the operator real time.
Intended operating mode. The model is one half of a tuned pair. The weights carry the de-risking and the domain SFT; the companion K3-Offsec-Engine carries methodology — progressive skill and tool turns, engagement state, and route receipts. The engine is strongly recommended, not decorative: the measured gains in output quality and thinking cost come from the pairing, not from the weights alone.
What it is not. Not a general-purpose chat model, not a drop-in production release, and not an unbounded model. The de-risking removes false refusals; it does not remove the operator's obligation to scope and supervise the work.
Why The Pairing Matters
The weights make the model answer. The engine makes the answers professional — and cheaper to produce.
Measured running the same engine build (0.4.0a1), five runs per arm at temperature 1.0:
| What was tested | Weights alone | Weights + engine |
|---|---|---|
| Attack planning (0–5) | 3.6, never perfect in 5 runs | 5.0, perfect in all 5 |
| Logic-bug discovery (0–5) | 3.5 | 3.8 — no measurable change |
| Thinking cost (reasoning tokens) | 100% | ~45% (−55%) |
| Answer completeness | terser | ~2× longer |
The pattern is consistent: where the bare model already performed well, the engine changed little; where methodology was the missing ingredient, it moved the result to the ceiling — and did it while spending roughly half the reasoning tokens. Less re-derivation, more answer.
Provenance. These figures were measured on a companion model with the same engine build, not on this checkpoint. They establish the engine's effect, not this model's absolute performance. Capability evaluation for this checkpoint is in progress (see status).
A legal-domain arm scored 65.7 → 66.0 of 69 (+0.5pp) on a 12-task exam — a null result, and reported as one: the bare model already scored 95%, leaving too little headroom to measure a difference. That test needs harder tasks before it can show anything.
What Was Done to the Model
Two modifications are applied to the upstream moonshotai/Kimi-K3.1-Pro checkpoint, and nothing else:
STAGE 01 · IN-PLACE WEIGHT EDIT
1. DWM Weight Edits (De-Risking)An in-place, targeted weight-space edit against the model's refusal direction. The edit subtracts the refusal direction from selected non-expert residual-stream writers while preserving the upstream architecture, expert routing, tokenizer, chat template, and quantization.
|
STAGE 02 · DOMAIN HARDENING
2. Blackfrost Cyber SFTA supervised fine-tune on a curated cybersecurity corpus (dual-use security knowledge, tool reasoning, and authorized-engagement conduct) to harden the model's security-domain competence and operational behavior.
|
Zero-Sidecar Weight Integrity: The modification is carried entirely in the weights — no runtime patch, serving overlay, or system-prompt bake is required to load or speak this checkpoint. The private method, recipe, and training data are not distributed in full; a reproducible recipe is published alongside the weights in
DWM-RECIPE/. An optional runtime engine is distributed in this repository's sidecar tree and is strongly recommended for security work, but the model loads and answers without it.
Interactive Widget · Modification Pipeline & Build Telemetry (2/4 Complete) — click to toggle
| Stage | Modification Step | Status | Technical Scope |
|---|---|---|---|
STAGE 01 · DWM-SUB |
DWM weight edits (de-risking) | ✓ COMPLETE |
An in-place, targeted weight-space edit against the model's refusal direction on selected non-expert residual-stream writers. |
STAGE 02 · CYBER-SFT |
Blackfrost Cyber SFT | ✓ COMPLETE |
Supervised fine-tune on a curated cybersecurity corpus to harden security-domain competence and operational conduct. |
STAGE 03 · VERIFY |
Post-build evaluation | ◐ PARTIAL |
False-refusal measured on a 443-prompt bench (a7.0: 302/443; R2: 267/443). Capability, agentic, and CyberGym evaluation in progress. |
STAGE 04 · PUBLISH |
Weights upload | ✓ COMPLETE |
Published — 96 SafeTensors shards (~1.42 TiB), build a7.0, with stock and R2 restore journals. |
Specifications (Inherited from Upstream)
Architecture, tokenizer, chat template, generation configuration, vision tower, and multimodal components are inherited from upstream unmodified in structure. Only the two disclosed weight modifications differ from the upstream checkpoint.
| Parameter / Subsystem | Specification |
|---|---|
| Upstream checkpoint | moonshotai/Kimi-K3.1-Pro |
| Architecture | KimiK31ForConditionalGeneration (kimi_k31) — hybrid-attention sparse MoE, omni (vision-language) |
| Attention mechanism | Hybrid attention architecture (upstream kimi_linear model type) |
| Layers | 93 |
| Hidden size | 7,168 |
| Routed experts | 896 routed · 16 experts per token · 2 shared experts |
| Attention heads | 96 (head_dim 128) |
| Vocabulary | 163,840 tokens |
| Context ceiling | 1,048,576 positions (upstream architectural maximum) |
| Modalities | Text and image input → Text output (Omni VLM) |
| Weight package (reference) | 96 SafeTensors shards · 1,561,217,377,280 indexed bytes (~1.42 TiB) |
| Build shipped | a7.0 — 186 axis-0 non-expert tensors, alpha 7.0, single pass |
| Alternative builds | Stock and residual-pass (R2) restore journals included as K31P_STOCK_BIN/ and K31P_R2_BIN/; see the swap procedure in K3-Offsec-Engine/AGENTS.md |
| Precision | Upstream tensor representation preserved as shipped; see upstream repository for precision details |
Cryptographic & Weight Lineage
moonshotai/Kimi-K3.1-Pro (Upstream Foundation Checkpoint)
│
├── [Stage 01] DWM De-Risking Weight Edits
│ └─ Refusal-direction subtraction on selected non-expert residual-stream writers
│ └─ Router gates, 896 MoE experts, vision tower & tokenizer preserved 1:1
│
└── [Stage 02] Blackfrost Cyber SFT
└─ Cybersecurity-domain supervised fine-tune (dual-use, tool reasoning, conduct)
│
▼
Blackfrost-Research/Kimi-K3.1-Pro-DERISKED ← THIS REPOSITORY
Blackfrost Research is responsible only for this derivative package. Moonshot AI is the upstream model developer and does not endorse this derivative.
Access: Manually Gated, Unlocked After Purchase
This repository is manually gated, and the checkpoint is unlocked only after purchase. To obtain access:
➜ Purchase access to this model — $599 — enter your Hugging Face username at checkout, and your access is granted automatically once payment is confirmed.
➜ Pay with USDC (Solana) — 599 USDC to C65NmgdrtppzXuE8wruPJaLE7NkBgawbCXjchdCkdSWH — include your Hugging Face username as the transfer memo, and click "Agree and request access" on this repo before sending. Access is granted once the transfer confirms (typically within the hour). Send only USDC on Solana; any other asset or network cannot be recovered.
|
STEP 01 · COMMERCIAL ARRANGEMENT Complete Purchase Agreement Complete the commercial purchase directly via the payment link above — $599 one-time, card accepted at checkout. |
STEP 02 · HUGGING FACE GATE Submit Repository Gate Request Submit your access request through the Hugging Face manual gate prompt on this repository page. |
STEP 03 · MANUAL UNLOCK Verification & Provisioning Access is granted manually after the purchase and organization identity are confirmed on our side. |
License terms for access requests are the upstream Kimi K3.1 Pro License terms (see License). Manual gating and purchase gating are access controls, not safety guarantees.
Verification and Evaluation Status
Qualification Progress: [███░░░░░░░░░░░░░░░░░] 1/6 Checks Complete
Interactive Widget · Verification & Qualification Gate Matrix — click to toggle
| Verification Check | Domain | Status | Acceptance & Audit Scope |
|---|---|---|---|
| Repository and card scope audit | AUDIT |
✓ COMPLETE |
Lineage disclosure, upstream license attribution, and non-endorsement boundaries verified. |
| Post-build weight-package validation | AUDIT |
✓ COMPLETE |
Shard integrity across 96 SafeTensors files (1,561,217,377,280 bytes) and tensor key parity verified against the published restore journals. |
| General capability evaluation | CAPABILITY |
⧗ PENDING |
Reasoning retention, instruction following, code synthesis, and multi-step tool execution. |
| Safety and behavioral evaluation | SAFETY |
◐ PARTIAL |
False-refusal rate measured on a 443-prompt bench (a7.0: 302/443; R2: 267/443). Benign-but-sensitive cybersecurity task completion and calibration profile in progress. |
| Red-line floor verification | SAFETY |
⧗ PENDING |
Confirmation that mandatory Blackfrost release red lines remain intact. |
| Long-context and vision qualification | QUALIFICATION |
⧗ PENDING |
1,048,576 context window stability and multimodal vision-tower alignment verification. |
No benchmark, throughput, refusal-rate, safety, or capability claim is made for this derivative at this time. Upstream results must not be treated as measured results for this checkpoint.
Infrastructure & Serving Reference (Post-Unlock)
Interactive Widget · Multi-Node VRAM Sizing & Serving Configurations (SGLang / vLLM) — click to expand
Full Upstream-Precision Reference Cluster
- Precision Profile:
Native Shipped Precision (96 Shards) - Memory Footprint:
~1.42 TiB indexed weights + KV cache headroom - Target Topology: 2× 8×H200 (141GB) or 2× 8×B200 NVLink Node Cluster
- Operational Note: Preserves exact upstream tensor representation across all 93 layers and 896 routed experts.
sglang.launch_server --model-path Blackfrost-Research/Kimi-K3.1-Pro-DERISKED --tp 16 --ep 16 --trust-remote-code --context-length 1048576
High-Density Research Assessment Node
- Precision Profile:
Multi-Node Tensor + Expert Parallelism (TP8/PP2) - Memory Footprint:
~1,600 GB aggregate cluster memory - Target Topology: 16× H100 80GB SXM5 (InfiniBand RoCEv2)
- Operational Note: Recommended for isolated air-gapped security evaluation and Cyber-Frost Harness validation.
vllm serve Blackfrost-Research/Kimi-K3.1-Pro-DERISKED --tensor-parallel-size 8 --pipeline-parallel-size 2 --enable-expert-parallel --max-model-len 262144
Companion Sidecar: K3-Offsec-Engine
The zero-sidecar guarantee above applies to the weights: the modification
is carried entirely in the checkpoint, with no runtime patches, serving
overlays, or system-prompt bakes required to load or speak the model. The
companion operation layer, K3-Offsec-Engine, is distributed in this
repository's K3-Offsec-Engine/ tree and is strongly
recommended for security work: the model answers without it, but the engine
is the intended operating mode for authorized tool-driven engagements.
When authorized security-tool work is intended, K3-Offsec-Engine serves as an engine-neutral OpenAI-compatible gateway in front of the private inference route. It owns progressive skill/MCP tool turns, keeps the raw model port private, and emits prompt-free route receipts. It does not alter weights, tokenizer files, or the serving template.
k3-offsec-engine serve-gateway \
--config config.kimi.toml \
--token-file /private/path/to/gateway-token \
--engine-profile kimi-compatible \
--host 127.0.0.1 --port 8915 \
--receipt-root /private/path/to/gateway-receipts
k3-offsec-engine probe-gateway \
--base-url http://127.0.0.1:8915/v1 \
--token-file /private/path/to/gateway-token \
--receipt-root /private/path/to/gateway-receipts
Verified against this model family on 2026-11-11: the route probe passed
(ok: true, progressive capability activation skill:target-triage, receipt
f98275ba1e9b694cc62df15506d8565ba39fd8d7efe16b15ef04717f4a487cc4). The
kimi-compatible engine profile handles the two endpoint quirks observed for
Kimi K3.1 Pro routes (it drops chat_template_kwargs, which strict routers
reject with HTTP 400, and repairs empty tool-call function.name values on
history replay). Follow the sidecar AGENTS.md deployment contract and treat
the pairing as promotion-gated: do not label the gateway production-promoted
for a model/engine pair until probe-gateway passes against that exact
served route — hosted-API results do not transfer to self-hosted SGLang or
vLLM endpoints, and vice versa.
Operational Scope: Intended Use vs. Limitations
Intended Use (Authorized Scope)
|
Strict Limitations & Operational Risks
|
License and Attribution
This derivative is distributed under the upstream Kimi K3.1 Pro License (Copyright © 2026 Moonshot AI), included as LICENSE when weights publish. That license grants use, copying, modification, distribution, sublicensing, sale, deployment, fine-tuning, and derivative works; this release is a derivative work under it.
See the moonshotai/Kimi-K3.1-Pro repository for upstream documentation and attribution.
The information on this page is provided as-is. For questions or responsible reporting, use this repository's Discussions.
Kimi-K3.1-Pro-DERISKED · © 2026 Blackfrost Softwares Corp. · blackfrostai.com
- Downloads last month
- 55
Model tree for Blackfrost-Research/Kimi-K3.1-Pro-DERISKED
Base model
moonshotai/Kimi-K3.1-Pro