Open-weight model
2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl
by Siddharth M. Bhatia siddharthmb/2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl
2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl is an open-weight model from Siddharth M. Bhatia. It has 4.6B parameters and a 8,192-token context. At 16-bit it needs about 11 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 41 downloads a month.
Sparse transcoder adapter for google/gemma-2-2b: a ReLU branch dec(relu(enc(x))) on every MLP of the frozen base model, trained in bridging mode so that base plus branch matches google/gemma-2-2b-it.
Runs On
What it takes to serve 2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl (4.6B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 9.2 GB | 11.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 4.6 GB | 5.5 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 2.3 GB | 2.7 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 12, 2026.
Model Card
Sparse transcoder adapter for google/gemma-2-2b: a ReLU branch dec(relu(enc(x))) on every MLP of the frozen base model, trained in bridging mode so that base plus branch matches google/gemma-2-2b-it. The branch's features are the fine-tuning, made sparse and inspectable (paper). Load it with the repo's model classes; the adapter branch is part of the checkpoint, and settranscodersenabled(False) gives the hybrid model (base MLPs with the reference model's attention and embeddings): Chat with it, or steer one feature by its cantor id (--steer CANTORID:STRENGTH): Attribution graphs of this adapter: uv run --extra viz python -m analysis.attribution.graphs.buildgraphs --adaptercheckpoint…
Excerpt from the card by Siddharth M. Bhatia.
Configuration
- Architecture
- Gemma2ForCausalLMWithTranscoder
- Context length (tokens)
- 8,192
- Layers
- 26
- Hidden size
- 2,304
- Feed-forward size
- 9,216
- Attention heads
- 8
- Key/value heads
- 4
- Head dimension
- 256
- Vocabulary size
- 256,000
- Sliding window (tokens)
- 4,096
- RoPE base
- 10000
- Model type
- gemma2
Identity and Version
- Repository
- siddharthmb/2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl
- Publisher
- Siddharth M. Bhatia
- Task
- Not stated by the source
- Modality
- Other
- Library
- transformers
- Parameters
- 4.6B parameters
- Languages
- Not stated by the source
- Revision
- a2fa85aabd58d58ca95124957dd68d329ebad249
- First published
- 2026-07-01
- Last updated
- 2026-10-11
Files and Weights
13 files, 9.2 GB in total. The weights are 2 files totalling 9.2 GB in safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00002.safetensors | Weights | 5.0 GB | 8cf0e891f5fa |
| model-00002-of-00002.safetensors | Weights | 4.2 GB | a630e6a069a9 |
| config.json | Configuration | 1.5 KB | — |
| generation_config.json | Configuration | 187 B | — |
| model.safetensors.index.json | Configuration | 33.0 KB | — |
| special_tokens_map.json | Configuration | 636 B | — |
| train_config.yaml | Configuration | 1.8 KB | — |
| README.md | Documentation | 5.6 KB | — |
| chat_template.jinja | Other | 591 B | — |
| .gitattributes | Repository | 1.6 KB | — |
| tokenizer.json | Tokenizer | 34.4 MB | 5f7eee611703 |
| tokenizer.model | Tokenizer | 4.2 MB | 61a7b147390c |
| tokenizer_config.json | Tokenizer | 46.4 KB | — |
License and Download
- License
- Not stated by the source
- Access
- Open weights, no gate
- Download size
- 9.2 GB
Released by Siddharth M. Bhatia through its official repository on Hugging Face.
Built From
- Derived from google/gemma-2-2b
- Described by arXiv:2602.20904
- Trained on (disclosed) science-of-finetuning/fineweb-1m-sample
- Trained on (disclosed) siddharthmb/2026.transcoder-adapters.lmsys-chat-1m-splits
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 9.2 GB |
| 16-bit | 9.2 GB |
| 8-bit | 4.6 GB |
| 4-bit | 2.3 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About 2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl
How much GPU memory does 2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl need?
About 11 GB at 16-bit and 2.7 GB at 4-bit: the weights (4.6B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run 2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
What is 2026.TA.gemma2_2b_huge_tc16384_decb_l1w0.0003_norm_sch_tarbb_lb2.0_ln1.0_dr500000_lr2e-04_bs8_sl's context length?
8,192 tokens, from the maximum position embeddings in its published configuration.