ABLAGENT-35B-A3B-FP8
BLACKFROST PRO SERIES · FLAGSHIP FP8 DEPLOYMENT EDITION
Controlled-access professional toolkit. This is the deployment-focused FP8 edition of
ABLAGENT-35B-A3B-BF16, packaged with a bounded standalone harness for professional model-engineering workflows. It is not presented as a general chat assistant, autonomous weight-editing system, or production- readiness guarantee.
Release status and contents
This repository is public and manually gated; every access request requires Blackfrost Research approval. It contains a complete standalone compressed-tensors FP8 checkpoint—not an adapter—plus tokenizer and multimodal processor assets, artifact checksums, a validated language-only deployment kit, and Blackfrost Ablagent Standalone Harness 1.0.0.
Blackfrost Pro Series releases combine a task-specific model artifact, a defined operating boundary, a harness or integration surface, deployment guidance, and a measured validation record. Tools, data access, write authority, and promotion decisions remain explicitly operator-governed.
| Pro Series standard | This release |
|---|---|
| Task-specific | Model-engineering evidence review and structured handoff |
| Toolkit-first | Compressed checkpoint, deployment kit, and bounded Ablagent harness |
| Private-deployment ready | Local-first operation and controlled endpoint support |
| Operator-governed | Read-only harness tools; no implicit execution or publication authority |
| Evidence-led | Conversion integrity and measured runtime checks are recorded separately |
Why this release exists
ABLAGENT-35B-A3B-FP8 reduces the weight payload of the canonical BF16 edition while preserving a deployable compressed-tensors representation for the validated SGLang path. It remains an Ablagent professional toolkit for inspecting artifacts, reconciling run evidence, comparing candidates, diagnosing incomplete experiments, and preparing clear operator handoffs.
Toolkit profile
| Series | Blackfrost Pro Series |
| Discipline | Model Engineering |
| Toolkit | Ablagent |
| Model | ABLAGENT-35B-A3B-FP8 |
| Primary workflow | Abliteration and Domain Weight Modification support |
| Operating mode | Evidence-focused, operator-directed, harness-bounded |
| Current tool boundary | Read-only workspace inspection |
| Delivery | Full FP8 checkpoint + deployment kit + standalone harness |
The harness can inspect an explicitly authorized workspace through bounded read_file, grep,
and glob tools. It does not expose a shell, modify weights, deploy services, publish artifacts,
or write into the workspace. Missing evidence remains missing rather than being treated as a pass.
Model specifications
| Item | Value |
|---|---|
| Architecture | MiniCPMV4_7ForConditionalGeneration |
| Language model | qwen3_5_moe_text |
| Model class | 35B total / approximately 3B active MoE |
| Transformer layers | 40 |
| Routed experts | 256 total / 8 selected per token |
| Released weight format | compressed-tensors FP8; FP8 E4M3 weights and dynamic FP8 activations |
| Weight shards | 8 indexed shards |
| Indexed weight payload | 37,878,423,648 bytes |
| Indexed tensors | 62,672 |
| Configured maximum context | 262,144 tokens |
| Delivery | Complete standalone checkpoint; no adapter required |
| Validated runtime | SGLang 0.5.20 CU130 on one NVIDIA B300 at TP1 |
| Pipeline | Image-text-to-text checkpoint; published SGLang kit is language-only |
Architecture and context values are checkpoint specifications. Runtime capacity is backend and hardware dependent. The included processor and vision payload were preserved, but multimodal quality was not re-benchmarked for this release.
Complete lineage
- Foundational checkpoint:
openbmb/MiniCPM-V-4.7-35B-A3B, immutable revision4c7623a9326e0f74313bbc7292d852a9e3018a70. - Blackfrost behaviorally modified derivative. The proprietary transformation process and its reconstruction materials are not distributed.
- Canonical BF16 parent:
Blackfrost-Research/ABLAGENT-35B-A3B-BF16, immutable revision9304f7634ceb66ba6a9a83698cac7982932b1ed2. - Direct FP8 deployment conversion using LLM Compressor
0.14.1.dev35+gdcd820dfat commitdcd820dfbad79b3191600f42278fc70e88c78e7b, compressed-tensors0.19.1a20261013, Transformers 5.19.0, and PyTorch 2.12.0+cu130.
The final conversion stage did not apply another behavioral transformation. Tokenizer, chat
template, and processor lineage comes from the pinned foundational checkpoint. The upstream and
this release declare Apache License 2.0; see
UPSTREAM-LICENSE-NOTICE.md.
Artifact verification and checksums
| Check | SHA-256 |
|---|---|
| Pinned upstream checkpoint digest | 46f798747cb038f1cee4179e81d0d86d303130531c8705ff8fa10d87a29d1a16 |
| Immediate BF16 parent artifact digest | af9b81adb22bbf3a94341466a9ea7e851d3941980c866b8dd7cf89f93af12361 |
Released config.json |
34415a71485f27677170725deb100a88977b6dfb813c7cbb53d3b62bc78b25e3 |
| Released weight index | daf0fc856c1ceee4e7a0341074220e39e7bbf8e048fae94ecaedccd862fae0ba |
| Artifact checksum manifest | 7cbf20c3af823350fe74ad2a40ff14282c4e973651a6599485314dacd160d215 |
Verify the downloaded runtime files from the repository root:
sha256sum --check ARTIFACT-CHECKSUMS.sha256
sha256sum --check DEPLOYMENT/CHECKSUMS.sha256
Behavioral and capability evaluation
- The output index contains 62,672 tensors with no missing, misplaced, or duplicate entries.
- 30,880 FP8 E4M3 tensors and their associated scale payloads passed structural and finite-value validation.
- The checkpoint loaded on one NVIDIA B300 at TP1 with a 262,144-token configured context.
/v1/modelsreturned exactlyABLAGENT-35B-A3B-FP8.- A deterministic minimal generation returned the correct result for
19 × 37. - The bundled harness passes 14 package, privacy, controller, and bounded-tool tests.
- These are conversion-integrity, structural, packaging, and liveness checks—not a comprehensive benchmark.
- No comprehensive multimodal, safety, long-context, or native tool-calling evaluation has been published for this derivative.
Prompt, template, and sampling notes
Use the included chat template and native OpenAI-compatible messages. The validated server enabled
thinking and used the qwen3 reasoning parser plus the qwen3_coder tool-call parser. Keep
reasoning_content separate from user-visible content and preserve the parser/template pairing.
The checkpoint generation configuration advertises temperature 1.0, top-k 20, and top-p 0.95. The minimal release smoke used temperature 0. Set sampling explicitly for every measured workload and provide enough output budget for both reasoning and the final response.
Deployment
DEPLOYMENT/ records the exact validated hardware class, runtime source commit,
container content ID, TP setting, launch flags, context, API shape, and smoke test. The backend must
support compressed-tensors FP8 W8A8 MoE. The launcher sets the clean API identity
ABLAGENT-35B-A3B-FP8; the smoke test refuses any other identity.
HARNESS/ contains the same Blackfrost Ablagent Standalone Harness used by the other
Ablagent releases. Configure and probe it only after the model endpoint passes the deployment-kit
smoke test:
cp HARNESS/config.35b.example.json HARNESS/config.local.json
HARNESS/bin/ablagent --config HARNESS/config.local.json --probe
Read HARNESS/docs/SECURITY.md,
HARNESS/docs/READINESS.md, and
HARNESS/docs/SERVER.md before broader use.
Limitations, security, license, and disclaimer
- The validated deployment path is language-only; it does not load the vision tower or multimodal projector.
- Configured 262K context is not a claim that every request length, concurrency level, or workload has been accepted.
- Quantization can change behavior relative to BF16; independently evaluate task quality.
- The model may refuse, comply incorrectly, hallucinate, or produce unsafe output. No claim of complete refusal removal, harmlessness, or production readiness is made.
- The current harness is a read-only evaluation surface, not a production automation controller.
- Operators are responsible for authorization, access control, prompt/data handling, tool permissions, output review, legal compliance, and independent evaluation.
- The checkpoint and bundled harness are provided under Apache License 2.0 with their respective notices. This is not a warranty of fitness for any purpose.
Do not redistribute access or mirror the gated weights without authorization. Report packaging, runtime, license, or security issues through this repository's Discussions tab without posting secrets, private prompts, or gated artifacts.
BLACKFROST PRO SERIES
Professional models. Defined workflows. Operator control.
ABLAGENT-35B-A3B-FP8 · © 2026 Blackfrost Softwares Corp.