Downloads · 30 days
64
9% of all-time downloads
Brooooooklyn/Gemma-4-26B-A4B-IT-UD-Q3_K_XL-mlx
Gemma-4-26B-A4B-IT-UD-Q3_K_XL-mlx is a text generation model from Brooooooklyn. Use it when you need the model to write or continue text. It is set up for mlx-node. The card lists the license as gemma.
3-bit affine quantization of google/gemma-4-26b-a4b-it for Apple Silicon, using the Unsloth Dynamic quantization strategy via mlx-node.
Downloads · 30 days
64
9% of all-time downloads
All-time downloads
710
Public
Parameters
25.8B
14.2 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors14.2 GB · 100%
How the weights are stored.
U3224.8B · 96%
From the Hugging Face model README
3-bit affine quantization of google/gemma-4-26b-a4b-it for Apple Silicon, using the Unsloth Dynamic quantization strategy via mlx-node.
| Original (BF16) | UD-Q3_K_XL (this model) | |
|---|---|---|
| Size | ~49 GB | 14 GB |
| Format | SafeTensors | SafeTensors |
| Precision | BF16 uniform | 3-bit affine + mixed bits + BF16 |
| FFN group size | — | 64 |
| Biases | — | yes |
| Repo | Bit budget | Size | Decode (tok/s) |
|---|---|---|---|
| Brooooooklyn/Gemma-4-26B-A4B-IT-UD-Q3_K_XL-mlx (this model) | 3-bit base | 14 GB | 60.6 |
| Brooooooklyn/Gemma-4-26B-A4B-IT-UD-MXFP4_K_XL-mlx | mxfp4 | 16 GB | 58.4 |
| Brooooooklyn/Gemma-4-26B-A4B-IT-UD-Q4_K_XL-mlx | 4-bit base | 17 GB | 58.6 |
| Brooooooklyn/Gemma-4-26B-A4B-IT-UD-NVFP4_K_XL-mlx | nvfp4 | 17 GB | 57.9 |
| Brooooooklyn/Gemma-4-26B-A4B-IT-UD-Q5_K_XL-mlx | 5-bit base | 20 GB | 50.3 |
| Brooooooklyn/Gemma-4-26B-A4B-IT-UD-Q6_K_XL-mlx | 6-bit base | 23 GB | 51.9 |
| Brooooooklyn/Gemma-4-26B-A4B-IT-UD-MXFP8_K_XL-mlx | mxfp8 | 26 GB | 49.8 |
| Brooooooklyn/Gemma-4-26B-A4B-IT-UD-Q8_K_XL-mlx | 8-bit base | 27 GB | 49.8 |
Benchmarked on Apple M3 Max 128GB via examples/lm.ts (best decode tok/s across turns 2–4, steady-state, capitals chat with reasoningEffort: 'low').
Note: No Q2 variant is published — Gemma-4-26B-A4B-IT has only ~4B active parameters per token, which is below the architectural redundancy needed for 2-bit quantization to remain coherent. Both unsloth and mixed_2_6 recipes produced gibberish at Q2 on this model.
Steady-state decode: 60.6 tok/s on Apple M3 Max 128GB (best of turns 2–4, examples/lm.ts capitals chat with reasoningEffort: 'low'). Decode is memory-bandwidth bound on Apple Silicon — fewer bytes per token directly translates to higher throughput. The MoE architecture activates only top-K of 128 experts per token (~4B active out of ~26B total), and the compiled C++ forward graph fuses the per-layer dispatch.
| Weight | Mode | Bits | Group | Rationale |
|---|---|---|---|---|
embed_tokens | 5-bit affine | 5 | 64 | Tied with lm_head (Gemma4 shares weights); affine-only loader |
self_attn.q_proj | 5-bit affine | 5 | 64 | AWQ-corrected via input_layernorm |
self_attn.k_proj | 5-bit affine | 5 | 64 | AWQ-corrected via input_layernorm |
self_attn.v_proj | 5-bit affine | 5 | 64 | AWQ-corrected via input_layernorm (only on full-attention layers) |
mlp.gate_proj | 3-bit affine | 3 | 64 | Shared dense MLP (top-level default) |
mlp.up_proj | 3-bit affine | 3 | 64 | Shared dense MLP (top-level default) |
mlp.down_proj | 4-bit affine | 4 | 64 | Shared dense MLP; "slightly more sensitive" (unsloth base+1) |
experts.switch_glu.gate_proj | 3-bit affine | 3 | 64 | MoE expert gate (per-expert across all 128); base bits (top-level default) |
experts.switch_glu.up_proj | 3-bit affine | 3 | 64 | MoE expert up (per-expert across all 128); base bits (top-level default) |
experts.switch_glu.down_proj | 4-bit affine | 4 | 64 | MoE expert down (per-expert across all 128 + routing); unsloth base+1 |
router.proj | 8-bit affine | 8 | 64 | MoE routing — low-bit noise breaks top-K expert selection |
self_attn.o_proj | bf16 | — | — | NOT AWQ-correctable; kept full-precision |
Built on Unsloth Dynamic 2.0 per-tensor KLD analysis. At --q-bits 3 the unsloth recipe assigns the base bits to MLP gate/up projections (the bulk of the parameter budget), base+1 to down_proj (slightly more sensitive), base+2 (snapped to a valid bit width) + AWQ pre-scaling to attention q/k/v projections, base+2 to embed_tokens, base+3 (capped/snapped) to the routing-critical paths, and keeps self_attn.o_proj as bf16 (AWQ-uncorrectable — its inputs come from the attention compute, not from a norm layer). The MoE router (router.proj) is forced to 8-bit affine to preserve top-K expert selection accuracy.
imatrix AWQ pre-scaling amplifies important weight channels and fuses inverse scales into preceding layer norms (zero inference overhead).
| Parameter | Value |
|---|---|
| Total parameters | ~26B (~4B active per token) |
| Hidden size | 2,816 |
| Layers | 30 (sliding-window attention) |
| Attention heads | 16 (8 KV heads, GQA 2:1) |
| Head dimension | 256 |
| Experts | 128 per MoE layer |
| MoE intermediate size | 704 |
| Vocab size | 262,144 |
| Max context | 262,144 tokens |
| Vision | yes (Gemma4ForConditionalGeneration) |
import { loadSession } from '@mlx-node/lm';
const session = await loadSession('./Gemma-4-26B-A4B-IT-UD-Q3_K_XL-mlx');
for await (const event of session.sendStream('Explain the MoE architecture in Gemma-4.', {
config: { maxNewTokens: 2048, temperature: 0.6, reasoningEffort: 'low' },
})) {
if (!event.done) process.stdout.write(event.text);
}
mlx convert \
-i gemma-4-26b-a4b-it \
-o Gemma-4-26B-A4B-IT-UD-Q3_K_XL-mlx \
-q --q-bits 3 --q-recipe unsloth \
--imatrix-path imatrix_unsloth.gguf
Gemma Terms of Use (inherited from base model).