smashburger-dev/kleinhirn-weights

kleinhirn weights

Converted weights for the kleinhirn engine. kleinhirn runs small encoder models in the browser on WebGPU, with a WASM-SIMD fallback on the CPU. These files are for that engine. They are not a Transformers checkpoint.

Each model sits in a folder named after its source repository. The f16 folder is for WebGPU with shader-f16, the f32 folder for WebGPU f32 and WASM.

Folder	Source model	Revision	License	Task	f16 MB	f32 MB
sentence-transformers/all-MiniLM-L6-v2/	sentence-transformers/all-MiniLM-L6-v2	1110a243fdf4	apache-2.0	embeddings	45.4	90.5
cardiffnlp/twitter-roberta-base-sentiment-latest/	cardiffnlp/twitter-roberta-base-sentiment-latest	3216a57f2a0d	cc-by-4.0	sequence-classification	249.9	499.2
cross-encoder/mmarco-mMiniLMv2-L12-H384-v1/	cross-encoder/mmarco-mMiniLMv2-L12-H384-v1	1427fd652930	apache-2.0	reranking	238.3	473.6
distilbert/distilbert-base-uncased-finetuned-sst-2-english/	distilbert/distilbert-base-uncased-finetuned-sst-2-english	714eb0fa89d2	apache-2.0	sequence-classification	134.2	268.1
protectai/deberta-v3-base-prompt-injection-v2/	protectai/deberta-v3-base-prompt-injection-v2	90c9989b1a34	apache-2.0	sequence-classification	388.3	775.3
ibm-granite/granite-embedding-small-english-r2/	ibm-granite/granite-embedding-small-english-r2	2ab6fa8ea2d6	apache-2.0	embeddings	95.9	191.2
fastino/gliner2.5-small-v1/	fastino/gliner2.5-small-v1	7e6f537f1033	apache-2.0	gliner2-classification	152.6	303.9
fastino/gliner2.5-base-v1/	fastino/gliner2.5-base-v1	b0c10b23313e	apache-2.0	gliner2-classification	390.0	778.8
fastino/gliner2.5-multi-v1/	fastino/gliner2.5-multi-v1	235cf92d6d43	apache-2.0	gliner2-classification	578.9	1155.2
SupersonicLabs/Julia-1/	SupersonicLabs/Julia-1	a85b127321d5	apache-2.0	julia-decision	293.3	581.9

Sizes are the sums of the files each manifest names (tokenizer, shards, words.bin), in 10^6 bytes.

Files

Each precision folder has:

• manifest.json: tensor layout, file sizes and sha256 of every file, and the source revision.
• tokenizer.bin: the upstream tokenizer.json in binary sections, decoded back to the identical object.
• shard-*.bin: the tensor data.
• words.bin: the word-embedding table, with a 32-bit checksum after each row.

The engine verifies every file against the sha256 in the manifest. With words: 'rows' it fetches only the rows of the tokens in a text by HTTP Range and checks each row's checksum. Such requests show this host which tokens a text contains.

Use

Point the kleinhirn engine at <org>/<name>/f16/manifest.json in this repository, at a pinned revision and not at main.

License and credit

Each model keeps its own license. See the license column above and the NOTICE in each model folder. The LICENSE file at the root is the Apache License 2.0 text.