LFM2.5-24B-A2B-DERISKED-FP8

ModelOpt FP8 release of the behaviorally modified LFM2.5-24B-A2B derivative for reasoning, multilingual chat, and tool-enabled workflows.

Release status: public and ungated. The standalone FP8 checkpoint, configuration, tokenizer, model card, account banner, license, and deployment kit are included.

Artifact contents

This is a standalone quantized checkpoint. It is not an adapter and does not require the BF16 parent at inference time.

Why this release exists

This release reduces the storage and serving footprint of the Blackfrost BF16 derivative while preserving its architecture, reasoning template, and tool-use interface. The weights are behaviorally modified. The proprietary transformation process and reconstruction materials are not distributed.

Model specifications

Complete lineage

1. Foundational parent: LiquidAI/LFM2.5-24B-A2B at revision 1769516ac588c4c558089c94f27d96b0dc67b183
2. Behavioral parent: Blackfrost-AI/LFM2.5-24B-A2B-DERISKED-BF16
3. Blackfrost stage: behaviorally modified derivative; the proprietary transformation process is not distributed
4. Quantization stage: ModelOpt FP8 export from the completed BF16 derivative
5. Public artifact: standalone quantized SafeTensors checkpoint

No upstream or BF16 benchmark number is presented as a post-quant measurement of this artifact.

Artifact verification and checksums

The completed FP8 artifact passed structural, numerical, and serving validation:

• all 3 expected weight shards are present;
• all 7,610 BF16 source tensor keys are retained;
• 7,402 eligible weights carry ModelOpt FP8 quantization payloads;
• the artifact contains 22,414 tensors totaling 24,239,130,448 tensor bytes;
• all 22,414 floating-point tensors scanned are finite;
• no zero-valued scale tensors were found;
• configuration, tokenizer, generation configuration, and chat-template assets are present;
• native MTP/NextN payload accounting is 0 → 0;
• vLLM auto-detected the modelopt loader, loaded the checkpoint at the 131,072-token context ceiling, exposed the OpenAI-compatible API, returned a separate reasoning field, and completed a deterministic generation canary on one NVIDIA B300.

The Hub revision records the complete checkpoint file inventory. File hashes are included in SHA256SUMS, and deployment-kit hashes are included in DEPLOYMENT/CHECKSUMS.sha256.

Measured behavioral and capability evaluation

The BF16 behavioral parent passed its narrow automatic preservation gate, with a relative negative-log-likelihood delta of -0.00305%, Jensen–Shannon divergence of 0.006415, and token top-1 agreement of 95.48% against its pinned parent on eight held-out prompts and 155 scored tokens.

Those distribution measurements describe the BF16 parent, not this quantized artifact. The FP8 artifact passed its post-quant structural, numerical, load, and deterministic generation canaries. A comprehensive post-quant academic, coding, multilingual, safety, refusal-behavior, and capability benchmark suite remains pending.

Prompt, reasoning, and sampling notes

Use the packaged chat template. It begins assistant turns in reasoning mode and supports structured tool definitions and tool-result reinjection. The upstream recommended defaults are temperature=0.1, top_k=50, and repetition_penalty=1.05.

Applications should validate their own system prompts, sampling settings, structured-output schemas, and tool policies. This release does not prescribe a default persona or application policy.

Deployment

The included DEPLOYMENT/ kit contains the exact validated runtime, launch flags, environment variables, smoke test, dependency provenance, hardware profile, and checksums for this artifact. Start with DEPLOYMENT/README.md.

Limitations

• Broad post-quant capability and behavioral evaluations remain pending.
• Quantization can change logits, output selection, and long-context behavior relative to BF16.
• The full configured context has not been exhaustively qualified under every serving topology.
• Tool calls are model-generated data and require application-side validation and permission controls.
• Outputs may be inaccurate, incomplete, biased, or fabricated. High-impact decisions require independent review.

Security responsibility

Deployers are responsible for authentication, authorization, rate limits, network isolation, logging, sandboxing, secrets handling, and tool-side validation. Treat retrieved content and tool output as untrusted input.

License

The parent model and this derivative are distributed under the LFM Open License v1.0. The full license text controls over this summary and includes commercial-use conditions. Quantization or redistribution does not replace or relax those terms.

Disclaimer

“DERISKED” is the repository name, not a safety certification. This research release is provided as-is, without warranty.

Reporting

Report reproducible issues through this repository's Hugging Face Discussions tab. Include the repository revision, runtime version or image digest, hardware class, serving flags, prompt/template settings, sampling parameters, and a minimal redacted reproduction.