bvolpato/kevala-packs

kevala packs

Pre-converted Q8 decoder packs and a BF16 vision companion for kevala, which runs decision models in the browser on WebGPU or WebAssembly. Loading a published pack skips in-browser conversion. Q8 packs download about half the bytes of the original checkpoints; the vision companion preserves BF16 weights.

file	model	from	size
laya-q8.kevala	Laya	convaiinnovations/laya at 1c5edc17a7acd8701df6fc341c0d179f1c62c982	479 MB
kev-0.8b-q8.kevala	Kev-0.8B	jaredpalmer/kev-0.8b at 54f4f8777356cd5bbbb6c6919c657f26e6f2f6d8, with its base Qwen/Qwen3.5-0.8B-Base at dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68	857 MB
kev-4b-q8.kevala	Kev-4B	jaredpalmer/kev-4b at 485ace8703592fcf405488b262449990824cfed1	4.76 GB
kev-9b-q8.kevala	Kev-9B	jaredpalmer/kev-9b at 2629c06a5aeb0feb3b9783bafed17ed8f39ecf5c	8.96 GB
semif-qwen3.5-0.8b-q8.kevala	SemIf Qwen3.5-0.8B	Qwen/Qwen3.5-0.8B at 2fc06364715b967f1860aea9cf38778875588b17	855 MB
semif-qwen3.5-2b-q8.kevala	SemIf Qwen3.5-2B	Qwen/Qwen3.5-2B at 15852e8c16360a2fea060d615a32b45270f8a8fc	2.13 GB
semif-qwen3.5-4b-q8.kevala	SemIf Qwen3.5-4B	Qwen/Qwen3.5-4B at 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a	4.75 GB
gemma-4-e2b-q8.kevala	Gemma 4 E2B	google/gemma-4-E2B-it at 3e22461f65e89153144f8adb70e3b8c2cc9845a7	5.217 GB
gemma-4-e4b-q8.kevala	Gemma 4 E4B	google/gemma-4-E4B-it at ee0ef6023621cff504d758262d4e04895a5af4a2	8.407 GB
gemma-4-e2b-vision-bf16.kevala	Gemma 4 E2B vision companion	google/gemma-4-E2B-it at 3e22461f65e89153144f8adb70e3b8c2cc9845a7	337.302 MB

The Laya, Kev, and SemIf packs were made with the kevala command line: weights are int8 with one f32 scale per 32 values (symmetric absmax); norms, biases, gates and decision heads stay f32; Kev's LoRA adapter is merged in f32 before quantization. The preferred command is kevala convert <checkpoint-dir> -o out.kevala; convert-kev, convert-semif, and convert-gemma remain compatibility aliases for the same validated path. The header of every pack records its source, revision, author, license and conversion. Against the upstream PyTorch code on the fixtures, Laya keeps 41/41 argmax agreement (largest probability difference 0.024) and Kev-0.8B 13/13 (0.010).

Sizes above use decimal units. Each new pack has an adjacent JSON manifest with its exact byte count, SHA-256, base checkpoint revision, and configuration. Kev-4B uses Qwen3.5-4B-Base at 1001bb4d826a52d1f399e183466143f4da7b741b; Kev-9B uses Qwen3.5-9B-Base at 68c46c4b3498877f3ef123c856ecfde50c39f404. These are text-only packs.

The Gemma 4 E2B and E4B packs are also text-only. They use the common kevala convert path (with convert-gemma retained as a compatibility alias) to extract the dense language model from Google's multimodal checkpoints and do not include the vision or audio towers. The sizes above are the converted text pack sizes.

Load gemma-4-e2b-vision-bf16.kevala alongside the matching E2B Q8 decoder to enable images on WebGPU. This companion contains the pinned vision tower and image projector, with BF16 matrix and positional weights, FP32 norms and clipping bounds, and FP32 arithmetic. It adds no quantization. Its exact size is 337,302,144 bytes and SHA-256 is 6556225c040a5b1d0bc8874a2833ce17cb7579d390f3225995c09ae589214d04. The adjacent JSON manifest records source identity, configuration, and validation provenance. E4B image loading and audio input are not implemented by this companion.

SemIf method

SemIf provides the direct-options-v1 prompt and scoring method used by the SemIf Qwen packs and adapted for the Gemma packs. These packs use frozen instruction weights, with no SemIf adapter or fine-tuning. One forward pass scores the next-token labels A-P; Kevala keeps only those output-head rows in f32 and applies a softmax over the declared options. It does not generate an answer or chain of thought. SemIf Qwen reuses evidence prefixes between questions; the initial Gemma backend runs each question independently.

The Qwen 4B model follows SemIf's reference choice. The 0.8B and 2B variants apply the same method to smaller models. All support up to 16 options per question. Their scores are conditional on the listed options and uncalibrated as decision confidence. These are Kevala conversions, not official SemIf checkpoints or a drop-in replacement for the SemIf Python API.

Gemma uses Google's instruction chat template with thinking disabled, then applies the same direct option readout. It scores the final hidden state against the declared label rows and applies Gemma's final logit soft cap before the option softmax.

Validation and requirements

The existing packs were checked on an RTX 5070 Ti (16 GB), Firefox 152.0.3, Ubuntu, using hardware WebGPU. These small fixture checks measure conversion fidelity, not task accuracy or a model ranking:

Pack	Reference precision	Matching decisions	Largest absolute probability difference
Kev-4B	BF16, LoRA merged in F32	13/13	0.00986
Kev-9B	F32 on CPU	5/5	0.00481
SemIf 0.8B	F32 on CUDA	12/12	0.03108
SemIf 2B	F32 on CUDA	12/12	0.02901
SemIf 4B	BF16 on CUDA	12/12	0.02006
Gemma 4 E2B IT	BF16 on CUDA	12/12	0.01907
Gemma 4 E4B IT	BF16 on CUDA	12/12	0.01209

The Gemma packs target WebGPU in supported browsers or the native 64-bit CPU CLI. Browser CPU and Node WebAssembly loading require packs under 2 GiB; use the native 64-bit CLI for these larger packs. WebGPU needs sufficient GPU memory and per-buffer limits. Large packs remain opt-in and are not downloaded by routine CI. Caching avoids later downloads, but loading still transfers weights to the GPU and needs working memory.

The E2B BF16 vision companion matches the pinned Transformers F32 reference with maximum absolute error 1.9908e-5 and relative L2 error 1.9567e-6, within a 1e-4 guard. Its isolated encoder benchmark improved from 901 ms to 99 ms median after the vision kernel and scheduling changes. These component measurements exclude the decoder and camera path; full webcam FPS and person/expression classification accuracy remain unverified. See the retained measurements.

Use a Kevala runtime that supports the pack's architecture. Older published packages can reject newer packs; the current source and hosted demo include the architectures listed here.

The Gemma backend accepts at most 4096 input tokens per question. Overlong requests fail instead of silently dropping evidence. This runtime limit is smaller than the upstream model context limits.

import { Kevala } from "kevala";
const kevala = await Kevala.load({ model: "laya" }); // fetches laya-q8.kevala from here

For Laya and Kev-0.8B, from: "checkpoint" converts the original weights in the browser. Other models require these packs. The native converter reads the checkpoint's Qwen2Tokenizer configuration and applies its normalization and added-token overlay, so ordinary conversion needs no Python or Transformers tokenizer materialization. Pass --tokenizer only for an explicit, verified override. The pinned conversion helper is tools/convert_models.py; it invokes the common native path. The commands that made and checked these packs, and the steps to publish new ones, are in docs/packs.md.

Credits and licenses

• Laya is by Nandakishor M, Convai Innovations, under Apache-2.0.
• Kev models are by Jared Palmer, under Apache-2.0. Their Qwen3.5 base models are by the Qwen team, under Apache-2.0. Kev's trained LoRA adapters and pointer heads are preserved in these conversions.
• Qwen3.5 instruction weights are by the Qwen team, under Apache-2.0.
• Gemma 4 E2B and E4B instruction weights are by Google DeepMind and are published under Apache-2.0; their pinned source revisions are listed above. The packs retain the source model metadata.
• SemIf's prompt and direct option-logit readout are MIT-licensed, copyright 2026 TheoLeeCJ. The complete notice is in THIRD_PARTY_NOTICES.

The Kevala tooling and pack metadata are distributed under Apache-2.0. These packs are derivative works of the listed source models and carry each source model's license and behavior limits: read the original model cards before relying on them.