README: add wllama-safe Q4_K_M + Q2_K variants (embed @ q5_K)

README.md

@@ -24,8 +24,9 @@ pipeline_tag: text-generation
 > use it and for complying with all applicable laws.
 
 GGUF quantizations of [`evalengine/unbound-e4b`](https://huggingface.co/evalengine/unbound-e4b)
-for on-device deployment via Ollama, llama.cpp, LM Studio, and similar
-desktop runtimes.
+for on-device deployment via Ollama, llama.cpp, LM Studio,
+[wllama](https://github.com/ngxson/wllama) (in-browser — see the wllama
+section below), and similar runtimes.
 
 Built by [Chromia](https://x.com/Chromia) and [Eval Engine](https://x.com/eval_engine).
 
@@ -35,20 +36,35 @@ All quants ship as **split multi-part GGUFs** (`*-00001-of-0000N.gguf` ...)
 so desktop runtimes can parallel-download chunks. Ollama, llama.cpp, and LM
 Studio auto-stitch on the first part — same UX as a single file.
 
-| Quant   | Parts | Total   | Largest part | wllama (browser) | Desktop (Ollama / llama.cpp / LM Studio) | Notes |
-|---------|-------|---------|--------------|------------------|------------------------------------------|-------|
-| Q2_K    | 4     | 4.08 GB | ~2.15 GB     | ❌                | ✅                                        | Smallest disk footprint; biggest quality drop |
-| Q3_K_M  | 4     | 4.49 GB | ~2.15 GB     | ❌                | ✅                                        | Modest size win over Q4 (embedding precision dominates) |
-| Q4_K_M  | 4     | 4.94 GB | ~2.15 GB     | ❌                | ✅                                        | **Recommended on-device default — best size/quality** |
-| Q6_K    | 5     | 5.75 GB | ~2.2 GB      | ❌                | ✅                                        | Higher fidelity |
-| Q8_0    | 6     | 7.43 GB | ~2.7 GB      | ❌                | ✅                                        | Highest fidelity |
-
-**Why no wllama:** E4B's `per_layer_token_embd` is a single atomic tensor
-that exceeds 2 GB in every quant we ship (the smallest one, Q4_K_M, lands
-at ~2.2 GB for that tensor). wllama's underlying browser ArrayBuffer caps at
-2 GB, so no split scheme can fit. For in-browser inference, use
-[`evalengine/unbound-e2b-GGUF`](https://huggingface.co/evalengine/unbound-e2b-GGUF)
-instead — its tensors are small enough that Q4_K_M and Q6_K both fit.
+### Desktop builds (Ollama / llama.cpp / LM Studio)
+
+These keep `per_layer_token_embd` at the llama.cpp default of Q6_K, which
+maximizes quality but pushes the largest split part above 2 GB — fine for
+desktop, won't load in browser.
+
+| Quant   | Parts | Total   | Largest part | Notes |
+|---------|-------|---------|--------------|-------|
+| Q2_K    | 4     | 4.08 GB | ~2.15 GB     | Smallest disk footprint; biggest quality drop |
+| Q3_K_M  | 4     | 4.49 GB | ~2.15 GB     | Modest size win over Q4 (embedding precision dominates total size) |
+| Q4_K_M  | 4     | 4.94 GB | ~2.15 GB     | **Recommended desktop default — best size/quality** |
+| Q6_K    | 5     | 5.75 GB | ~2.2 GB      | Higher fidelity |
+| Q8_0    | 6     | 7.43 GB | ~2.7 GB      | Highest fidelity |
+
+### Browser builds (wllama)
+
+E4B's `per_layer_token_embd` is a single 2.82-billion-value tensor; at the
+default Q6_K precision it lands at ~2.2 GB, just over the browser
+ArrayBuffer cap. These variants force the embedding tensors to `q5_K`,
+which shrinks the largest part below 2 GB at near-zero quality cost.
+
+| Quant variant       | Parts | Total   | Largest part | wllama | Notes |
+|---------------------|-------|---------|--------------|--------|-------|
+| **Q4_K_M-wllama**   | 4     | 4.51 GB | **1848 MB**  | ✅      | **Recommended browser default** — layers @ Q4_K_M, embed @ q5_K |
+| **Q2_K-wllama**     | 4     | 3.69 GB | **1848 MB**  | ✅      | Smallest browser-loadable build — layers @ Q2_K, embed @ q5_K |
+
+The wllama files use the same split-multi-part naming, so Ollama / llama.cpp
+will auto-stitch them too if you prefer the smaller embed quant for any
+reason.
 
 ## Recommended sampling
 
@@ -76,6 +92,21 @@ ollama run  hf.co/evalengine/unbound-e4b-GGUF
 ./llama-cli -m unbound-e4b-Q4_K_M-00001-of-00004.gguf -p "your prompt"
 ```
 
+## Run in the browser (wllama)
+
+[wllama](https://github.com/ngxson/wllama) is a WebAssembly port of llama.cpp
+that runs entirely in the browser. Use one of the wllama-safe variants
+above:
+
+```js
+import { Wllama } from '@wllama/wllama';
+const wllama = new Wllama(/* … */);
+await wllama.loadModelFromHF(
+  'evalengine/unbound-e4b-GGUF',
+  'unbound-e4b-Q4_K_M-wllama-00001-of-00004.gguf'  // wllama follows the chain
+);
+```
+
 ## About the base
 
 See [`evalengine/unbound-e4b`](https://huggingface.co/evalengine/unbound-e4b)