Primary use stated or inferred cautiously from official documentation.
TML-d24-varloop-research
Research checkpoint released alongside Inkling 2 Mini on November 20. It allows the number of passes through reused middle layers to vary at inference. It retains 26.4 billion physical parameters and accepts text and images; it is not simply a quantization of Mini.
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
Limits and integration
| API ID | thinkingmachines/TML-d24-varloop-research |
|---|---|
| Model type | Multimodal · Code · Text and reasoning |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosted |
| Release | 20 NOV 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image |
| Exit | Text |
| Context window | 1M tokens |
| maximum output | Text; separate output maximum unpublished |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face |
Study how quality changes when the same weights receive one, two, three, four or five passes.
Distributed as weights only: not available through an API or Tinker. Each additional pass adds computation, not new physical parameters. Use at least 32 GB of memory and compatible runtimes; do not attribute Mini’s scores to this checkpoint.
What it can do
Orientation
Study how quality changes when the same weights receive one, two, three, four or five passes.
Context
1M tokens · Text; separate output maximum unpublished
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
No verified official API tariff. Self-hosting requires memory, compute and maintenance; downloadable weights do not make inference free.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
How to read the results
At three passes, the official table reports 47.4% on DeepSWE v1.1 and 84.2% on GPQA Diamond. The pass count is retained because results change with it. Maxima from different configurations are not combined.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| DeepSWE v1.1 (varloop 3 passes) | 47.4 % | mini-swe-agent v2.4.6 | View source ↗ |
| GPQA Diamond (varloop 3 passes) | 84.2 % | Published score | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
Chronology
thinkingmachines/TML-d24-varloop-research
Version added to Inferama's verified catalog.