# qwen3.5-4l-vocab40k-en-ko-headless by Junsung Kim
Source: https://savrn.com/models/qwen3-5-4l-vocab40k-en-ko-headless
Markdown alternate of the page above; the site index is https://savrn.com/llms.txt

---

## Runs On

What it takes to serve qwen3.5-4l-vocab40k-en-ko-headless (121M 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 | 0.2 GB | 0.3 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 |
| 8-bit | 0.1 GB | 0.1 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 |
| 4-bit | 0.1 GB | 0.1 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/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](https://savrn.com/ai-index/pricing/gpus), read Oct 7, 2026.

[qwen3.5-4l-vocab40k-en-ko-headless on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/qwen3-5-4l-vocab40k-en-ko-headless/gpus)

## Model Card

By Junsung Kim, published under apache-2.0, revision e906337bd193.

### Qwen3.5 4L Headless Backbone (40k EN/KO)

A headless four-layer text backbone distilled from Qwen3.5-0.8B, with a vocabulary cut to English and Korean. It has no language-model head, no classification head, no labels and no thresholds. You attach a head and train it.

Korean translation of this page: README_KO.md.

This card is the record of how it was made. It walks the whole path — prompting the original model, turning it into a classifier, removing layers, changing precision, cutting the vocabulary, and cutting it again by rule — and shows what each step measured. We are not arguing that this model is better than anything. We built it, we measured it beside other models, and where we have a guess about why a number moved we say it is a guess.

[Read the full model card (7,839 words)](https://savrn.com/models/qwen3-5-4l-vocab40k-en-ko-headless/card)

## Configuration

Architecture

Qwen3_5TextModel

Context length (tokens)

262,144

Layers

4

Hidden size

1,024

Feed-forward size

3,584

Attention heads

8

Key/value heads

2

Head dimension

256

Vocabulary size

39,866

Model type

qwen3_5_text

## Identity and Version

Repository

mp-juuuns/qwen3.5-4l-vocab40k-en-ko-headless

Publisher

Junsung Kim

Task

Feature extraction

Modality

Text

Library

transformers

Parameters

121M parameters

Languages

en, ko

Revision

e906337bd193061825bcb3e24b34829a80e1779f

First published

2026-07-31

Last updated

2026-09-25

## Files and Weights

377 files, 1.7 GB in total. The weights are 5 files totalling 1.6 GB in gguf, safetensors.

Weights5 files · 1.6 GB

Configuration231 files · 5.0 MB

Tokenizer13 files · 34.6 MB

Documentation26 files · 220.9 KB

Other99 files · 11.3 MB

Repository3 files · 3.3 KB

Every file

| File | Type | Size | SHA-256 |
| --- | --- | --- | --- |
| docker/bundle/model/classifier-q8_0.gguf | Weights | 129.8 MB | e92fb36bf62b |
| model.safetensors | Weights | 241.3 MB | 694ce8325365 |
| models/semeval-propaganda/model.safetensors | Weights | 421.8 MB | 0352da6e4717 |
| models/semeval-propaganda/seeds/seed42/model.safetensors | Weights | 421.8 MB | 2db5f4be391d |
| models/semeval-propaganda/seeds/seed43/model.safetensors | Weights | 421.8 MB | 81a56d497df7 |
| benchmark/cost_isolation/MEASUREMENTS.json | Configuration | 9.7 KB | — |
| benchmark/cost_isolation/encoder_matched_settings.json | Configuration | 1.9 KB | — |
| benchmark/cost_isolation/matrix_eager_and_compile.json | Configuration | 332 B | — |
| benchmark/cost_isolation/matrix_fla_and_fla_compile.json | Configuration | 329 B | — |
| benchmark/cost_isolation/measure_compile_numerics.py | Configuration | 1.4 KB | — |
| benchmark/cost_isolation/measure_cost_matrix.py | Configuration | 3.0 KB | — |
| benchmark/cost_isolation/measure_encoders.py | Configuration | 4.4 KB | — |
| benchmark/cost_isolation/measure_stride.py | Configuration | 2.7 KB | — |
| benchmark/cost_probe/TOKENIZATION_COST.json | Configuration | 1.1 KB | — |
| benchmark/cost_probe/WINDOW_COST.json | Configuration | 1.6 KB | — |
| benchmark/cost_probe/measure_tokenization_cost.py | Configuration | 2.8 KB | — |
| benchmark/cost_probe/measure_window_cost.py | Configuration | 3.9 KB | — |
| benchmark/figures/annotate_01.py | Configuration | 1.5 KB | — |
| benchmark/figures/fig10.py | Configuration | 6.6 KB | — |
| benchmark/figures/make_cost_figures.py | Configuration | 8.6 KB | — |
| benchmark/figures/make_cost_isolation_figure.py | Configuration | 6.4 KB | — |
| benchmark/figures/make_more_figures.py | Configuration | 13.2 KB | — |
| benchmark/figures/make_readout_figures.py | Configuration | 26.3 KB | — |
| benchmark/figures/make_structure_figures.py | Configuration | 11.7 KB | — |
| benchmark/reports/existing_specialized_separate_lineage-seed41.json | Configuration | 115.8 KB | — |
| benchmark/reports/existing_specialized_separate_lineage-seed42.json | Configuration | 115.8 KB | — |
| benchmark/reports/existing_specialized_separate_lineage-seed43.json | Configuration | 116.0 KB | — |
| benchmark/reports/structural_control-seed41.json | Configuration | 115.6 KB | — |
| benchmark/reports/structural_control-seed42.json | Configuration | 115.2 KB | — |
| benchmark/reports/structural_control-seed43.json | Configuration | 115.7 KB | — |
| benchmark/reports/structural_copy_control-seed41.json | Configuration | 115.6 KB | — |
| benchmark/reports/structural_copy_control-seed42.json | Configuration | 115.2 KB | — |
| benchmark/reports/structural_copy_control-seed43.json | Configuration | 115.7 KB | — |
| benchmark/reports/task_agnostic-seed41.json | Configuration | 115.9 KB | — |
| benchmark/reports/task_agnostic-seed42.json | Configuration | 115.7 KB | — |
| benchmark/reports/task_agnostic-seed43.json | Configuration | 115.1 KB | — |
| benchmark/reports/task_agnostic_base-seed41.json | Configuration | 115.9 KB | — |
| benchmark/reports/task_agnostic_base-seed42.json | Configuration | 115.7 KB | — |
| benchmark/reports/task_agnostic_base-seed43.json | Configuration | 115.1 KB | — |
| config.json | Configuration | 2.5 KB | — |
| distillation/configs/task_agnostic_base.yaml | Configuration | 357 B | — |
| distillation/examples/labels.json | Configuration | 36 B | — |
| distillation/qwen35_distill/__init__.py | Configuration | 175 B | — |
| distillation/qwen35_distill/checkpoint.py | Configuration | 13.2 KB | — |
| distillation/qwen35_distill/classification.py | Configuration | 14.6 KB | — |
| distillation/qwen35_distill/cli.py | Configuration | 8.4 KB | — |
| distillation/qwen35_distill/layer_maps.py | Configuration | 3.1 KB | — |
| distillation/qwen35_distill/losses.py | Configuration | 3.8 KB | — |
| distillation/qwen35_distill/schema.py | Configuration | 5.6 KB | — |
| distillation/qwen35_distill/training.py | Configuration | 11.7 KB | — |
| docker-train/compose.arm64.yaml | Configuration | 510 B | — |
| docker-train/compose.yaml | Configuration | 2.5 KB | — |
| docker-train/config/showcase.yaml | Configuration | 765 B | — |
| docker-train/config/train.yaml | Configuration | 305 B | — |
| docker-train/docs/provenance/converter-files.json | Configuration | 15.7 KB | — |
| docker-train/docs/provenance/runtime-files.json | Configuration | 263.7 KB | — |
| docker-train/docs/provenance/semeval2020-license-review.json | Configuration | 1.1 KB | — |
| docker-train/examples/data/showcase/manifest.json | Configuration | 3.1 KB | — |
| docker-train/make_base.py | Configuration | 3.6 KB | — |
| docker-train/native/patch_runtime.py | Configuration | 17.0 KB | — |
| docker-train/provenance/converter.json | Configuration | 15.7 KB | — |
| docker-train/provenance/dependency-fixtures.json | Configuration | 579 B | — |
| docker-train/qwen35_deploy/__init__.py | Configuration | 257 B | — |
| docker-train/qwen35_deploy/__main__.py | Configuration | 1.8 KB | — |
| docker-train/qwen35_deploy/contract.py | Configuration | 6.8 KB | — |
| docker-train/qwen35_deploy/data.py | Configuration | 5.5 KB | — |
| docker-train/qwen35_deploy/export.py | Configuration | 8.7 KB | — |
| docker-train/qwen35_deploy/metrics.py | Configuration | 1.4 KB | — |
| docker-train/qwen35_deploy/native.py | Configuration | 5.5 KB | — |
| docker-train/qwen35_deploy/service.py | Configuration | 21.1 KB | — |
| docker-train/qwen35_deploy/showcase.py | Configuration | 4.1 KB | — |
| docker-train/qwen35_deploy/telemetry.py | Configuration | 5.9 KB | — |
| docker-train/qwen35_deploy/train.py | Configuration | 10.6 KB | — |
| docker-train/tests/test_contract.py | Configuration | 3.9 KB | — |
| docker-train/tests/test_release.py | Configuration | 2.7 KB | — |
| docker-train/tests/test_showcase.py | Configuration | 5.2 KB | — |
| docker-train/tests/test_training.py | Configuration | 3.5 KB | — |
| docker-train/tools/audit_release.py | Configuration | 5.6 KB | — |
| docker-train/tools/build_native.py | Configuration | 4.2 KB | — |
| docker-train/tools/check_arm.py | Configuration | 2.7 KB | — |
| docker-train/tools/check_container.py | Configuration | 10.4 KB | — |
| docker-train/tools/check_http.py | Configuration | 2.5 KB | — |
| docker-train/tools/check_native_heads.py | Configuration | 3.3 KB | — |
| docker-train/tools/check_showcase.py | Configuration | 9.9 KB | — |
| docker-train/tools/collect_licenses.py | Configuration | 1.7 KB | — |
| docker-train/tools/collect_package_notices.py | Configuration | 1.5 KB | — |
| docker-train/tools/package_base.py | Configuration | 4.1 KB | — |
| docker-train/tools/package_release.py | Configuration | 4.8 KB | — |
| docker-train/tools/prepare_showcase.py | Configuration | 5.8 KB | — |
| docker-train/tools/smoke_local.py | Configuration | 3.2 KB | — |
| docker-train/validation/arm-container-pytorch.json | Configuration | 183 B | — |
| docker-train/validation/arm-functional.json | Configuration | 1.6 KB | — |
| docker-train/validation/browser.json | Configuration | 538 B | — |
| docker-train/validation/calibration-arm64.json | Configuration | 39.5 KB | — |
| docker-train/validation/calibration-cpu.json | Configuration | 40.1 KB | — |
| docker-train/validation/calibration-cuda.json | Configuration | 41.3 KB | — |
| docker-train/validation/container-amd64-cpu.json | Configuration | 2.6 KB | — |
| docker-train/validation/container-amd64-cuda.json | Configuration | 2.8 KB | — |
| docker-train/validation/container-arm64-cpu.json | Configuration | 2.7 KB | — |
| docker-train/validation/container-arm64-inference.json | Configuration | 535 B | — |
| docker-train/validation/http.json | Configuration | 3.4 KB | — |
| docker-train/validation/image-layers-amd64.json | Configuration | 615 B | — |
| docker-train/validation/image-layers-arm64.json | Configuration | 443 B | — |
| docker-train/validation/local-summary.json | Configuration | 1.6 KB | — |
| docker-train/validation/native-heads.json | Configuration | 1.6 KB | — |
| docker-train/validation/showcase-image-layers.json | Configuration | 3.4 KB | — |
| docker-train/validation/showcase-reload-alignment.json | Configuration | 726 B | — |
| docker-train/validation/showcase.json | Configuration | 1.9 KB | — |
| docker-train/validation/training-cpu.json | Configuration | 314 B | — |
| docker-train/validation/training-cuda.json | Configuration | 315 B | — |
| docker-train/vendor/converter/conversion/__init__.py | Configuration | 14.2 KB | — |
| docker-train/vendor/converter/conversion/afmoe.py | Configuration | 3.2 KB | — |
| docker-train/vendor/converter/conversion/arctic.py | Configuration | 6.8 KB | — |
| docker-train/vendor/converter/conversion/baichuan.py | Configuration | 2.5 KB | — |
| docker-train/vendor/converter/conversion/bailingmoe.py | Configuration | 9.3 KB | — |
| docker-train/vendor/converter/conversion/base.py | Configuration | 136.2 KB | — |
| docker-train/vendor/converter/conversion/bert.py | Configuration | 26.9 KB | — |
| docker-train/vendor/converter/conversion/bitnet.py | Configuration | 1.8 KB | — |
| docker-train/vendor/converter/conversion/bloom.py | Configuration | 2.9 KB | — |
| docker-train/vendor/converter/conversion/chameleon.py | Configuration | 2.4 KB | — |
| docker-train/vendor/converter/conversion/chatglm.py | Configuration | 8.4 KB | — |
| docker-train/vendor/converter/conversion/codeshell.py | Configuration | 973 B | — |
| docker-train/vendor/converter/conversion/cogvlm.py | Configuration | 933 B | — |
| docker-train/vendor/converter/conversion/command_r.py | Configuration | 7.9 KB | — |
| docker-train/vendor/converter/conversion/dbrx.py | Configuration | 3.6 KB | — |
| docker-train/vendor/converter/conversion/deci.py | Configuration | 9.3 KB | — |
| docker-train/vendor/converter/conversion/deepseek.py | Configuration | 36.3 KB | — |
| docker-train/vendor/converter/conversion/dots1.py | Configuration | 1.2 KB | — |
| docker-train/vendor/converter/conversion/dotsocr.py | Configuration | 2.1 KB | — |
| docker-train/vendor/converter/conversion/dream.py | Configuration | 2.9 KB | — |
| docker-train/vendor/converter/conversion/ernie.py | Configuration | 8.6 KB | — |
| docker-train/vendor/converter/conversion/exaone.py | Configuration | 14.8 KB | — |
| docker-train/vendor/converter/conversion/falcon.py | Configuration | 2.7 KB | — |
| docker-train/vendor/converter/conversion/falcon_h1.py | Configuration | 5.5 KB | — |
| docker-train/vendor/converter/conversion/gemma.py | Configuration | 47.8 KB | — |
| docker-train/vendor/converter/conversion/glm.py | Configuration | 12.5 KB | — |
| docker-train/vendor/converter/conversion/gpt2.py | Configuration | 3.2 KB | — |
| docker-train/vendor/converter/conversion/gpt_oss.py | Configuration | 5.9 KB | — |
| docker-train/vendor/converter/conversion/gptneox.py | Configuration | 2.9 KB | — |
| docker-train/vendor/converter/conversion/granite.py | Configuration | 24.6 KB | — |
| docker-train/vendor/converter/conversion/grok.py | Configuration | 5.2 KB | — |
| docker-train/vendor/converter/conversion/grovemoe.py | Configuration | 4.8 KB | — |
| docker-train/vendor/converter/conversion/hunyuan.py | Configuration | 17.5 KB | — |
| docker-train/vendor/converter/conversion/internlm.py | Configuration | 10.9 KB | — |
| docker-train/vendor/converter/conversion/internvl.py | Configuration | 4.5 KB | — |
| docker-train/vendor/converter/conversion/jais.py | Configuration | 4.0 KB | — |
| docker-train/vendor/converter/conversion/jamba.py | Configuration | 5.0 KB | — |
| docker-train/vendor/converter/conversion/januspro.py | Configuration | 4.3 KB | — |
| docker-train/vendor/converter/conversion/kimi_linear.py | Configuration | 11.6 KB | — |
| docker-train/vendor/converter/conversion/kimivl.py | Configuration | 7.0 KB | — |
| docker-train/vendor/converter/conversion/lfm2.py | Configuration | 10.3 KB | — |
| docker-train/vendor/converter/conversion/lighton_ocr.py | Configuration | 882 B | — |
| docker-train/vendor/converter/conversion/llada.py | Configuration | 7.0 KB | — |
| docker-train/vendor/converter/conversion/llama.py | Configuration | 20.0 KB | — |
| docker-train/vendor/converter/conversion/llama4.py | Configuration | 1.6 KB | — |
| docker-train/vendor/converter/conversion/llava.py | Configuration | 6.0 KB | — |
| docker-train/vendor/converter/conversion/maincoder.py | Configuration | 409 B | — |
| docker-train/vendor/converter/conversion/mamba.py | Configuration | 9.6 KB | — |
| docker-train/vendor/converter/conversion/mellum.py | Configuration | 2.5 KB | — |
| docker-train/vendor/converter/conversion/mimo.py | Configuration | 13.5 KB | — |
| docker-train/vendor/converter/conversion/minicpm.py | Configuration | 8.5 KB | — |
| docker-train/vendor/converter/conversion/minimax.py | Configuration | 1.9 KB | — |
| docker-train/vendor/converter/conversion/mistral.py | Configuration | 9.5 KB | — |
| docker-train/vendor/converter/conversion/mistral3.py | Configuration | 2.3 KB | — |
| docker-train/vendor/converter/conversion/mpt.py | Configuration | 2.1 KB | — |
| docker-train/vendor/converter/conversion/nemotron.py | Configuration | 18.1 KB | — |
| docker-train/vendor/converter/conversion/olmo.py | Configuration | 4.3 KB | — |
| docker-train/vendor/converter/conversion/openelm.py | Configuration | 3.7 KB | — |
| docker-train/vendor/converter/conversion/orion.py | Configuration | 1.7 KB | — |
| docker-train/vendor/converter/conversion/pangu.py | Configuration | 1.8 KB | — |
| docker-train/vendor/converter/conversion/phi.py | Configuration | 18.0 KB | — |
| docker-train/vendor/converter/conversion/pixtral.py | Configuration | 1.4 KB | — |
| docker-train/vendor/converter/conversion/plamo.py | Configuration | 8.5 KB | — |
| docker-train/vendor/converter/conversion/plm.py | Configuration | 786 B | — |
| docker-train/vendor/converter/conversion/qwen.py | Configuration | 32.7 KB | — |
| docker-train/vendor/converter/conversion/qwen3vl.py | Configuration | 16.3 KB | — |
| docker-train/vendor/converter/conversion/qwenvl.py | Configuration | 9.0 KB | — |
| docker-train/vendor/converter/conversion/refact.py | Configuration | 3.2 KB | — |
| docker-train/vendor/converter/conversion/rwkv.py | Configuration | 14.2 KB | — |
| docker-train/vendor/converter/conversion/sarashina2.py | Configuration | 958 B | — |
| docker-train/vendor/converter/conversion/smallthinker.py | Configuration | 3.5 KB | — |
| docker-train/vendor/converter/conversion/smolvlm.py | Configuration | 2.1 KB | — |
| docker-train/vendor/converter/conversion/stablelm.py | Configuration | 4.0 KB | — |
| docker-train/vendor/converter/conversion/starcoder.py | Configuration | 885 B | — |
| docker-train/vendor/converter/conversion/step3.py | Configuration | 16.5 KB | — |
| docker-train/vendor/converter/conversion/t5.py | Configuration | 14.1 KB | — |
| docker-train/vendor/converter/conversion/talkie.py | Configuration | 2.1 KB | — |
| docker-train/vendor/converter/conversion/ultravox.py | Configuration | 8.8 KB | — |
| docker-train/vendor/converter/conversion/xverse.py | Configuration | 3.9 KB | — |
| docker-train/vendor/converter/conversion/youtuvl.py | Configuration | 2.9 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/__init__.py | Configuration | 219 B | — |
| docker-train/vendor/converter/gguf-py/gguf/constants.py | Configuration | 181.1 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/gguf.py | Configuration | 478 B | — |
| docker-train/vendor/converter/gguf-py/gguf/gguf_reader.py | Configuration | 15.1 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/gguf_writer.py | Configuration | 60.8 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/lazy.py | Configuration | 9.5 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/metadata.py | Configuration | 39.3 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/quants.py | Configuration | 64.9 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/scripts/gguf_convert_endian.py | Configuration | 7.0 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/scripts/gguf_dump.py | Configuration | 23.1 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/scripts/gguf_editor_gui.py | Configuration | 64.4 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/scripts/gguf_hash.py | Configuration | 3.7 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/scripts/gguf_new_metadata.py | Configuration | 10.2 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/scripts/gguf_set_metadata.py | Configuration | 4.1 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/tensor_mapping.py | Configuration | 115.3 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/utility.py | Configuration | 13.5 KB | — |
| docker-train/vendor/converter/gguf-py/gguf/vocab.py | Configuration | 42.4 KB | — |
| docker-train/vendor/manifest.json | Configuration | 263.7 KB | — |
| docker/benchmark_reference.json | Configuration | 2.3 KB | — |
| docker/bundle/manifest.json | Configuration | 264.2 KB | — |
| docker/bundle/model.json | Configuration | 1.1 KB | — |
| docker/compose.yaml | Configuration | 495 B | — |
| docker/prepare_bundle.py | Configuration | 4.9 KB | — |
| docker/server.py | Configuration | 26.3 KB | — |
| docker/telemetry.py | Configuration | 5.9 KB | — |
| docker/test_server.py | Configuration | 8.7 KB | — |
| finalization_manifest.json | Configuration | 1.6 KB | — |
| models/semeval-propaganda/benchmark/final4l_aggregate.json | Configuration | 10.9 KB | — |
| models/semeval-propaganda/benchmark/seed41.json | Configuration | 115.8 KB | — |
| models/semeval-propaganda/benchmark/seed42.json | Configuration | 115.8 KB | — |
| models/semeval-propaganda/benchmark/seed43.json | Configuration | 116.0 KB | — |
| models/semeval-propaganda/classify.py | Configuration | 8.1 KB | — |
| models/semeval-propaganda/config.json | Configuration | 3.4 KB | — |
| models/semeval-propaganda/labels.json | Configuration | 891 B | — |
| models/semeval-propaganda/release_manifest.json | Configuration | 30.9 KB | — |
| models/semeval-propaganda/seeds/seed42/config.json | Configuration | 3.4 KB | — |
| models/semeval-propaganda/seeds/seed43/config.json | Configuration | 3.4 KB | — |
| models/semeval-propaganda/thresholds.json | Configuration | 2.6 KB | — |
| provenance/base/finalization_manifest.json | Configuration | 1.6 KB | — |
| provenance/base/training_manifest.json | Configuration | 3.6 KB | — |
| provenance/task_agnostic_stages/24to8.json | Configuration | 3.6 KB | — |
| provenance/task_agnostic_stages/6to4.json | Configuration | 3.6 KB | — |
| provenance/task_agnostic_stages/8to6.json | Configuration | 3.6 KB | — |
| root_manifest.json | Configuration | 3.8 KB | — |
| training_manifest.json | Configuration | 3.6 KB | — |
| v128k_remap_manifest.json | Configuration | 1.1 KB | — |
| LICENSE | Documentation | 11.4 KB | — |
| README.md | Documentation | 53.6 KB | — |
| README_KO.md | Documentation | 63.5 KB | — |
| benchmark/BENCHMARK_CARD.md | Documentation | 3.2 KB | — |
| benchmark/figures/README.md | Documentation | 1.8 KB | — |
| benchmark/figures/cross_task/README.md | Documentation | 868 B | — |
| distillation/README.md | Documentation | 3.4 KB | — |
| docker-train/README.md | Documentation | 4.9 KB | — |
| docker-train/README_KO.md | Documentation | 11.4 KB | — |
| docker-train/VALIDATION_KO.md | Documentation | 11.3 KB | — |
| docker-train/docs/REFERENCES_KO.md | Documentation | 4.4 KB | — |
| docker-train/docs/RELEASE_POLICY_KO.md | Documentation | 4.1 KB | — |
| docker-train/docs/SHOWCASE_KO.md | Documentation | 2.9 KB | — |
| docker-train/docs/THIRD_PARTY_NOTICES.md | Documentation | 3.8 KB | — |
| docker-train/docs/licenses/llama.cpp-LICENSE | Documentation | 1.1 KB | — |
| docker-train/docs/licenses/llama.cpp-converter-LICENSE | Documentation | 1.1 KB | — |
| docker-train/docs/licenses/runtime/LICENSE | Documentation | 1.1 KB | — |
| docker-train/docs/licenses/runtime/cmake/license.cmake | Documentation | 1.4 KB | — |
| docker-train/examples/data/showcase/NOTICE.md | Documentation | 2.9 KB | — |
| docker-train/vendor/converter/LICENSE | Documentation | 1.1 KB | — |
| docker/README.md | Documentation | 5.3 KB | — |
| docker/README_KO.md | Documentation | 5.8 KB | — |
| docs/RELEASE_CONTRACT.md | Documentation | 5.4 KB | — |
| models/semeval-propaganda/LICENSE | Documentation | 11.4 KB | — |
| models/semeval-propaganda/README.md | Documentation | 1.4 KB | — |
| runtime/README.md | Documentation | 2.4 KB | — |
| SHA256SUMS | Other | 41.1 KB | — |
| benchmark/arm_summary.csv | Other | 1.7 KB | — |
| benchmark/compression_ladder.csv | Other | 827 B | — |
| benchmark/cost_isolation/build_causal_conv1d.sh | Other | 2.0 KB | — |
| benchmark/cost_isolation/stride_sweep.txt | Other | 225 B | — |
| benchmark/figures/01_overall_transfer_quality.csv | Other | 500 B | — |
| benchmark/figures/01_overall_transfer_quality.pdf | Other | 18.1 KB | — |
| benchmark/figures/01_overall_transfer_quality.png | Other | 82.1 KB | — |
| benchmark/figures/01_quality_vs_latency.png | Other | 83.6 KB | — |
| benchmark/figures/02_per_label_transfer_metrics.csv | Other | 1.2 KB | — |
| benchmark/figures/02_per_label_transfer_metrics.pdf | Other | 21.4 KB | — |
| benchmark/figures/02_per_label_transfer_metrics.png | Other | 138.3 KB | 5b558738b405 |
| benchmark/figures/02_top_arms_overlap.png | Other | 108.1 KB | bdc9cb91e84f |
| benchmark/figures/03_depth_quality_and_cost.png | Other | 91.2 KB | — |
| benchmark/figures/03_quality_efficiency.csv | Other | 448 B | — |
| benchmark/figures/03_quality_efficiency.pdf | Other | 21.9 KB | — |
| benchmark/figures/03_quality_efficiency.png | Other | 99.4 KB | — |
| benchmark/figures/04_compression_ladder.csv | Other | 352 B | — |
| benchmark/figures/04_compression_ladder.pdf | Other | 19.5 KB | — |
| benchmark/figures/04_compression_ladder.png | Other | 120.8 KB | 9f6c1d6fdefd |
| benchmark/figures/04_vocabulary_cut_curve.png | Other | 90.5 KB | — |
| benchmark/figures/05_taskblind_grid.png | Other | 144.1 KB | 422e77215c23 |
| benchmark/figures/06_stage_path.png | Other | 134.5 KB | 4599ab058b0b |
| benchmark/figures/07_runtime_specialization.png | Other | 123.3 KB | c9c09f7f91a7 |
| benchmark/figures/08_long_context_window.png | Other | 130.9 KB | 9519e737b99e |
| benchmark/figures/10_root_on_every_metric.png | Other | 181.6 KB | 66d828782589 |
| benchmark/figures/11_final_test_calibration_to_test.png | Other | 198.8 KB | 71aa42d84f71 |
| benchmark/figures/12_final_test_per_label.png | Other | 189.7 KB | cfb3071db407 |
| benchmark/figures/13_final_test_seed_spread.png | Other | 156.8 KB | 4a9704a2f64b |
| benchmark/figures/14_field_quality_vs_cost.png | Other | 216.7 KB | c097205e0424 |
| benchmark/figures/15_field_metric_spans.png | Other | 129.7 KB | 67f3a002bfeb |
| benchmark/figures/16_field_depth_effect.png | Other | 205.8 KB | 5da006b4d603 |
| benchmark/figures/17_field_resolution_floor.png | Other | 314.1 KB | dc621e266a69 |
| benchmark/figures/18_vocabulary_cost_inversion.png | Other | 285.5 KB | 8941a6f6bcf3 |
| benchmark/figures/19_window_cost.png | Other | 220.1 KB | 261f2fddb02e |
| benchmark/figures/20_where_the_parameters_live.png | Other | 178.6 KB | 87c6e141f8f2 |
| benchmark/figures/21_ladder_loss.png | Other | 165.3 KB | 8dc8781f83f6 |
| benchmark/figures/22_threshold_and_overprediction.png | Other | 244.2 KB | b08436d88b8a |
| benchmark/figures/23_what_costs_the_time.png | Other | 207.1 KB | d169de992c48 |
| benchmark/figures/cross_task/05_why_4l_edge_tradeoff.csv | Other | 5.4 KB | — |
| benchmark/figures/cross_task/05_why_4l_edge_tradeoff.pdf | Other | 27.2 KB | — |
| benchmark/figures/cross_task/05_why_4l_edge_tradeoff.png | Other | 179.0 KB | cdc0a8a98cd1 |
| benchmark/figures/cross_task/06_malint_4l_quality.csv | Other | 2.1 KB | — |
| benchmark/figures/cross_task/06_malint_4l_quality.pdf | Other | 26.4 KB | — |
| benchmark/figures/cross_task/06_malint_4l_quality.png | Other | 181.4 KB | 2c5b22415d31 |
| benchmark/figures/cross_task/07_malint_quality_resource.csv | Other | 1.8 KB | — |
| benchmark/figures/cross_task/07_malint_quality_resource.pdf | Other | 29.3 KB | — |
| benchmark/figures/cross_task/07_malint_quality_resource.png | Other | 163.6 KB | 1cb13cc695f3 |
| benchmark/figures/v128k_compression.csv | Other | 198 B | — |
| benchmark/figures/v128k_quality_efficiency.csv | Other | 370 B | — |
| benchmark/figures/v128k_quality_storage.png | Other | 73.4 KB | — |
| benchmark/figures/v128k_vocabulary_size.png | Other | 52.8 KB | — |
| benchmark/full/61_arm_bf16.csv | Other | 4.9 KB | — |
| benchmark/full/taskblind_grid.csv | Other | 399 B | — |
| benchmark/full/vocabulary_cuts_same_host.csv | Other | 583 B | — |
| benchmark/per_label_metrics.csv | Other | 13.0 KB | — |
| benchmark/per_label_summary.csv | Other | 7.0 KB | — |
| benchmark/resource_metrics.csv | Other | 1.6 KB | — |
| benchmark/seed_metrics.csv | Other | 1.2 KB | — |
| chat_template.jinja | Other | 6.7 KB | — |
| distillation/examples/multilabel_train.jsonl | Other | 186 B | — |
| distillation/examples/single_label_train.jsonl | Other | 171 B | — |
| distillation/examples/unlabeled.jsonl | Other | 183 B | — |
| distillation/pyproject.toml | Other | 487 B | — |
| docker-train/Dockerfile | Other | 3.6 KB | — |
| docker-train/docs/licenses/SemEval-CC-BY-4.0.txt | Other | 18.7 KB | — |
| docker-train/examples/data/calibration.jsonl | Other | 191 B | — |
| docker-train/examples/data/showcase/calibration.jsonl | Other | 4.0 KB | — |
| docker-train/examples/data/showcase/train.jsonl | Other | 9.2 KB | — |
| docker-train/examples/data/spans-calibration.jsonl | Other | 203 B | — |
| docker-train/examples/data/spans-train.jsonl | Other | 244 B | — |
| docker-train/examples/data/train.jsonl | Other | 303 B | — |
| docker-train/native/gdn_fused.inc | Other | 1.7 KB | — |
| docker-train/native/q8_kernel.inc | Other | 3.1 KB | — |
| docker-train/native/q8_neon_kernel.inc | Other | 2.2 KB | — |
| docker-train/native/q8_packed_kernel.inc | Other | 8.1 KB | — |
| docker-train/native/q8_zmm_packed.inc | Other | 2.2 KB | — |
| docker-train/native/quantize.cpp | Other | 650 B | — |
| docker-train/qwen35_deploy/app.js | Other | 12.7 KB | — |
| docker-train/qwen35_deploy/index.html | Other | 10.7 KB | — |
| docker-train/requirements-inference.lock | Other | 334 B | — |
| docker-train/requirements-training.lock | Other | 245 B | — |
| docker-train/tools/setup_validation_host.sh | Other | 3.4 KB | — |
| docker-train/vendor/native_live_classifier.cpp | Other | 5.5 KB | — |
| docker-train/vendor/runtime_source.tar.gz | Other | 3.0 MB | 7fdb374e511d |
| docker/Dockerfile | Other | 416 B | — |
| docker/app.js | Other | 13.0 KB | — |
| docker/bundle/native_live_classifier.cpp | Other | 5.5 KB | — |
| docker/bundle/runtime_source.tar.gz | Other | 3.0 MB | 7fdb374e511d |
| docker/index.html | Other | 10.2 KB | — |
| docker/requirements.lock | Other | 323 B | — |
| docker/start.sh | Other | 2.1 KB | — |
| models/semeval-propaganda/chat_template.jinja | Other | 6.7 KB | — |
| models/semeval-propaganda/requirements.txt | Other | 54 B | — |
| models/semeval-propaganda/seeds/seed42/chat_template.jinja | Other | 6.7 KB | — |
| models/semeval-propaganda/seeds/seed43/chat_template.jinja | Other | 6.7 KB | — |
| requirements.txt | Other | 49 B | — |
| runtime/qwen35-classifier.cpp | Other | 6.7 KB | — |
| runtime/qwen35.cpp | Other | 40.9 KB | — |
| .gitattributes | Repository | 2.7 KB | — |
| docker-train/.dockerignore | Repository | 408 B | — |
| docker/.dockerignore | Repository | 134 B | — |
| benchmark/figures/09_tokenizer_cost.png | Tokenizer | 72.8 KB | — |
| docker-train/vendor/converter/conversion/wavtokenizer.py | Tokenizer | 1.7 KB | — |
| docker/bundle/model/tokenizer.json | Tokenizer | 2.9 MB | — |
| merges.txt | Tokenizer | 381.7 KB | — |
| models/semeval-propaganda/seeds/seed42/tokenizer.json | Tokenizer | 9.2 MB | dd2b27bbd4f5 |
| models/semeval-propaganda/seeds/seed42/tokenizer_config.json | Tokenizer | 1.3 KB | — |
| models/semeval-propaganda/seeds/seed43/tokenizer.json | Tokenizer | 9.2 MB | dd2b27bbd4f5 |
| models/semeval-propaganda/seeds/seed43/tokenizer_config.json | Tokenizer | 1.3 KB | — |
| models/semeval-propaganda/tokenizer.json | Tokenizer | 9.2 MB | dd2b27bbd4f5 |
| models/semeval-propaganda/tokenizer_config.json | Tokenizer | 1.3 KB | — |
| tokenizer.json | Tokenizer | 2.9 MB | — |
| tokenizer_config.json | Tokenizer | 1.1 KB | — |
| vocab.json | Tokenizer | 731.3 KB | — |

## License and Download

License

apache-2.0

Access

Open weights, no gate

Download size

1.6 GB

[Download from Junsung Kim](https://huggingface.co/mp-juuuns/qwen3.5-4l-vocab40k-en-ko-headless)

Released by Junsung Kim through its official repository on Hugging Face. [Read the license](https://www.apache.org/licenses/LICENSE-2.0).

## Built From

- Derived from [Qwen/Qwen3.5-0.8B](https://savrn.com/models/qwen3-5-0-8b)
- Trained on (disclosed) [Salesforce/wikitext](https://savrn.com/datasets/wikitext)

## Memory Requirements

| Precision | Weights in memory |
| --- | --- |
| As published | 1.6 GB |
| 16-bit | 0.2 GB |
| 8-bit | 0.1 GB |
| 4-bit | 0.1 GB |

Weights only, from the published parameter count; the key-value cache and runtime add to this.

## Questions About qwen3.5-4l-vocab40k-en-ko-headless

### How much GPU memory does qwen3.5-4l-vocab40k-en-ko-headless need?

About 0.3 GB at 16-bit and 0.1 GB at 4-bit: the weights (121M parameters) plus a working margin. A long context needs more.

### What is the cheapest GPU to run qwen3.5-4l-vocab40k-en-ko-headless 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.

### Can I use qwen3.5-4l-vocab40k-en-ko-headless commercially?

Yes. qwen3.5-4l-vocab40k-en-ko-headless is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

### What is qwen3.5-4l-vocab40k-en-ko-headless's context length?

262,144 tokens, from the maximum position embeddings in its published configuration.

## Similar Models

Model · Feature extraction

### [tiny-audio-granite-qwen-top4](https://savrn.com/models/tiny-audio-granite-qwen-top4)

[Alex Kroman](https://savrn.com/model-publishers/mazesmazes)

This is the model card of a transformers model that has been pushed on the Hub. Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

Open weights 122M parameters transformers

[View model](https://savrn.com/models/tiny-audio-granite-qwen-top4)

Model · Feature extraction

### [bge-base-en-v1.5](https://savrn.com/models/bge-base-en-v1-5)

[Beijing Academy of Artificial Intelligence](https://savrn.com/model-publishers/baai)

For more details please refer to our Github: FlagEmbedding. If you are looking for a model that supports more languages, longer texts, and other retrieval methods, you can try using bge-m3. FlagEmbedding focuses on retrieval-augmented LLMs, consisting of the following projects currently: - 1/30/2024: Release BGE-M3, a new member to BGE model series! M3 stands for Multi-linguality (100+ languages), Multi-granularities (input length up to 8192), Multi-Functionality (unification of dense, lexical, multi-vec/colbert retrieval). It is the first embedding model which supports all three retrieval methods, achieving new SOTA on multi-lingual (MIRACL) and cross-lingual (MKQA) benchmarks. Technical…

Open weights mit 109M parameters 512 tokens sentence-transformers

[View model](https://savrn.com/models/bge-base-en-v1-5)

Model · Feature extraction

### [bge-base-en](https://savrn.com/models/bge-base-en)

[Beijing Academy of Artificial Intelligence](https://savrn.com/model-publishers/baai)

Recommend switching to newest BAAI/bge-base-en-v1.5, which has more reasonable similarity distribution and same method of usage. More details please refer to our Github: FlagEmbedding. FlagEmbedding can map any text to a low-dimensional dense vector which can be used for tasks like retrieval, classification, clustering, or semantic search. And it also can be used in vector databases for LLMs. Updates - 10/12/2023: Release LLM-Embedder, a unified embedding model to support diverse retrieval augmentation needs for LLMs. Paper:fire: - 09/15/2023: The technical report of BGE has been released - 09/15/2023: The masive training data of BGE has been released - 09/12/2023: New models: - 09/07/2023…

Open weights mit 109M parameters 512 tokens transformers

[View model](https://savrn.com/models/bge-base-en)

Model · Feature extraction

### [SapBERT-from-PubMedBERT-fulltext](https://savrn.com/models/sapbert-from-pubmedbert-fulltext)

[Language Technology Lab @University of Cambridge](https://savrn.com/model-publishers/cambridgeltl)

[news] A cross-lingual extension of SapBERT will appear in the main onference of ACL 2021! [news] SapBERT will appear in the conference proceedings of NAACL 2021! SapBERT by Liu et al. (2020). Trained with UMLS 2020AA (English only), using microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext as the base model. The input should be a string of biomedical entity names, e.g., "covid infection" or "Hydroxychloroquine". The [CLS] embedding of the last layer is regarded as the output. The following script converts a list of strings (entity names) into embeddings. For more details about training and eval, see SapBERT github repo.

Open weights apache-2.0 109M parameters 512 tokens transformers

[View model](https://savrn.com/models/sapbert-from-pubmedbert-fulltext)

Model · Feature extraction

### [PhaGen](https://savrn.com/models/phagen)

[Tim Trump](https://savrn.com/model-publishers/tim419)

PhaGen-0404-step12800 is the frozen paper model: 152,421,601 parameters, three hierarchical stages with dimensions 512/256/196, maximum token canvas 131,072, and a nine-token character-level DNA tokenizer. This repository contains the standalone inference weights, tokenizer, model configuration and custom code. Weights are the selected April 2026 run's global step 12,800, not the final saved checkpoint or a later training implementation. Use PyTorch 2.5, Transformers 4.54.1, einops 0.8, beartype 0.22, accelerate and safetensors. The PhaGen source repository also provides an installable phagen CLI and an agent skill. Pin a Hub commit hash with revision for repeatable loading. Custom model…

Open weights 152M parameters transformers

[View model](https://savrn.com/models/phagen)

Model · Feature extraction

### [jina-embeddings-v2-base-code](https://savrn.com/models/jina-embeddings-v2-base-code)

[Jina AI](https://savrn.com/model-publishers/jinaai)

The easiest way to starting using jina-embeddings-v2-base-code is to use Jina AI's Embedding API. jina-embeddings-v2-base-code is an multilingual embedding model speaks English and 30 widely used programming languages. Same as other jina-embeddings-v2 series, it supports 8192 sequence length. jina-embeddings-v2-base-code is based on a Bert architecture (JinaBert) that supports the symmetric bidirectional variant of ALiBi to allow longer sequence length. The backbone jina-bert-v2-base-code is pretrained on the github-code dataset. The model is further trained on Jina AI's collection of more than 150 millions of coding question answer and docstring source code pairs. These pairs were obtained…

Open weights apache-2.0 161M parameters 8,192 tokens sentence-transformers

[View model](https://savrn.com/models/jina-embeddings-v2-base-code)

## Junsung Kim

[All models and datasets](https://savrn.com/model-publishers/mp-juuuns)

## Versions

- [e906337bd193](https://savrn.com/models/qwen3-5-4l-vocab40k-en-ko-headless/versions/e906337bd193) · current 2026-09-25

## Explore More

- [All feature extraction models](https://savrn.com/models/tasks/feature-extraction)
- [All models under apache-2.0](https://savrn.com/models/licenses/apache-2-0)
- [Model comparisons](https://savrn.com/models/comparisons)
- [The model directory](https://savrn.com/models)
- [Open model prices by host](https://savrn.com/ai-index/pricing/open-models)

## Source

- Repository metadata, read 2026-09-25.
- [Hugging Face record](https://huggingface.co/mp-juuuns/qwen3.5-4l-vocab40k-en-ko-headless)
- [How the hub is built](https://savrn.com/model-hub/methodology)