INKLING-2-MINI-DERISKED
A behaviorally modified derivative of the Inkling-2-Mini loop transformer for controlled research and internal evaluation.
PRIVATE PREVIEW: Weights, deployment kit, and DWM recipe included. Read
deployment-kit/AGENT_INSTRUCTIONS.mdcompletely before serving.
Release status
| Item | Status |
|---|---|
| Language model | Weight-level behavioral modification validated |
| MTP / speculative component | Preserved and validated end-to-end |
| Deployment kit | deployment-kit/ (agent instructions + validated vendor serve config) |
| DWM recipe | DWM_RECIPE/ (scripts, hard32 battery, capture evidence, runbook) |
| Modification recipe | Proprietary and intentionally not distributed |
| Comprehensive derivative benchmark | In progress; initial 443-row refusal observations below |
Why this release exists
This build reduces false refusals on ordinary, safe requests while retaining a genuine-refusal floor on clearly harmful requests. It provides a behaviorally modified operating point of Inkling-2-Mini for controlled research and internal evaluation. This is the first derivative Blackfrost has produced on a looped-transformer (repeated-layer) architecture; directions were captured and applied per-block with the loop pass structure intact.
Model specifications
| Property | Value |
|---|---|
| Architecture | Inkling-2-Mini loop transformer (layer-looped causal language model) |
| Total parameters | ~17.6B (measured from published weight shard headers) |
| Effective active parameters | ~5.4B per token at the vendor's three-pass loop configuration |
| Weight format | Native vendor-quantized safetensors, unmodified quantization |
| Context | Upstream 262,144-token claim; serving validated at that window on the operator runtime |
| Concurrency validated | 8 parallel generation streams during evaluation |
| Speculative decoding | DSpark draft sampling, 7 speculative tokens, MTP preserved |
| License | Per the upstream Thinking Machines release terms |
The upstream context specification is not a claim that every downstream deployment was validated at the full window.
Lineage
thinkingmachines/Inkling-2-Mini
└── INKLING-2-MINI — DERISKED
| Stage | Provenance |
|---|---|
| Foundation | thinkingmachines/Inkling-2-Mini |
| Derivative stage | Weight-level behavioral modification; proprietary process not distributed |
| Quantization | Unchanged from upstream (native vendor quantization) |
| Runtime | vLLM with DSpark speculative decoding, per the vendor serving recipe |
Tokenizer, configuration, generation defaults, and the MTP module follow the upstream release byte-for-byte except for the modified target weight shard identified by checksum below.
Artifact verification
| Artifact | Bytes | SHA-256 |
|---|---|---|
Derivative weight shard (model-00001-of-00001.safetensors) |
19,526,019,168 | b1156b5d53ddabf927a8e3f209eebc5da610d2b90ed8c14a6a945b78080bef21 |
MTP module (mtp.safetensors), byte-identical to upstream |
635,801,416 | 7c3b651317f01f73bcf4e1dceba66ef8b7e1a1eac6bc171e882e690a4b10a1d2 |
s2-capture/directions.npz (48, 2304) float32 — 48 direction sites: layers 9–18, each with 4 loop visits + metadata |
406,708 | see DWM_RECIPE/RUNBOOK.md |
s2-capture/raw-capture.npz — 64 prompts × hidden states |
79,118,677 | see DWM_RECIPE/RUNBOOK.md |
The MTP module is byte-identical to upstream; MTP is always preserved after modification.
Behavioral and capability evaluation
Hand-adjudicated 443-row refusal observations, bare prompts, 32k answer budget, zero API errors:
| Observation | Result | Rate |
|---|---|---|
| Harm-arm compliance (advbench + strongreject) | 266/295 | 90.2% |
| Harm-arm real refusals | 29/295 | 9.8% |
| False refusals on safe prompts (xstest) | 0/75 | 0.0% |
These are adjudicated operational observations on this operator's runtime, not a controlled cross-model benchmark. Regex-flagged rows were hand-reviewed; false positives were excluded. Upstream benchmark scores do not transfer to this modified artifact.
Alpha sweep (50-prompt smoke set, bare, 32k budget): answer rate rose monotonically with edit scale (alpha 1.5 → 2.5: 21 → 38 answered; refusals 15 → 2) with zero errors and no coherence loss at the selected operating point.
Prompting, templates, and tools
The validated serving path uses the vendor chat template with the inkling reasoning parser. Reasoning is returned in a dedicated reasoning field. The runtime does not bundle an application-specific tool executor; applications remain responsible for tool schema handling, execution, and result injection.
Deployment
Card-only language removed. The validated configuration uses the vendor's vLLM image with DSpark speculative decoding (num_speculative_tokens=7, probabilistic draft sampling) at the full upstream context window. See deployment-kit/ for agent instructions and the validated serve command; see DWM_RECIPE/ for the reproduction runbook.
Limitations and security responsibility
- Behavioral modification was applied to a loop transformer; unusual loop-dependent behaviors should be re-validated per deployment.
- Evaluation was one runtime, one serving configuration, and a bounded prompt set.
- Treat model output as untrusted data. Validate generated code, tool arguments, URLs, file paths, and commands before execution.
- The model is not an authorization, policy-enforcement, or security boundary.
- Operators are responsible for access control, rate limiting, audit logging, data handling, and compliance in their deployment environment.
License and disclaimer
Released subject to the upstream Inkling-2-Mini model terms and notices. Review the upstream license before redistribution or deployment.
This research artifact is provided without warranties. Users are responsible for validating fitness, accuracy, security, and legal compliance for their use case.
For reproducible artifact issues, use the repository's issue-tracking channel without posting secrets or private infrastructure details.
- Downloads last month
- 70