fuutott/Qwen3.8-Flash-Next-MLX-t5-imatrix-q3down-ple8

⚠️ Runtime requirement. This checkpoint only runs on the Qwen48 oMLX fork: the ternary expert banks need its native Bonsai Metal kernel. Stock oMLX, mlx-lm and mlx-vlm cannot load it. See Running it.

Qwen3.8-Flash-Next for a 48 GB Mac (MLX, ternary experts + Q3 down, imatrix)

A mixed-precision MLX checkpoint of Qwen/Qwen3.8-Flash-Next that runs on a 48 GB Apple Silicon Mac with the Qwen48 oMLX fork. The 360 GB source model is brought down to about 35.6 GiB resident by storing the routed experts' gate/up projections as ternary weights (oMLX's Bonsai T5 packing, ~1.875 bits/weight), the routed down projections as affine Q3, and the 53.6 GiB PLE n-gram embedding table as Q8 that is memory-mapped from SSD instead of loaded. All low-bit tensors are fitted with the Unsloth importance matrix for this model.

Measured on an M3 Max 48 GB: 38 tok/s decode, 210–380 tok/s prefill, 64K-token chat requests completed, and a 129K-token prefill verified by oMLX's automatic context sizing. Quality sits between the previous ternary builds and Unsloth's Q4KXL GGUF, which does not fit this machine (numbers below). This is an experimental research build, not an official Qwen or oMLX release.

What is in the checkpoint

Every affine tensor whose input channels could be mapped onto the GGUF importance matrix is fitted with it (routed and shared experts, attention, DeltaNet, hyper-connection projections), including the DeltaNet out_proj, whose imatrix entry llama.cpp stores in a tiled V-head order that the converter undoes. The T5 packing stores five ternary digits per byte with a BF16 scale per group of 128; it is a packing format, not the T5 language-model family, and it is not Tencent AngelSlim's STQ1_0 layout, although the weighted-least-squares fit was inspired by it.

Architecture is unchanged from the source: qwen4_exp (hyper-connections, Gated DeltaNet and QSA attention layers, 512-expert MoE with top-10 routing, one PLE n-gram embedding layer, vision tower). It is not Qwen3.5; loaders that treat it as Qwen3.5 will fail.

Measured on the target machine

MacBook Pro M3 Max, 48 GB unified memory, macOS 27.0, fork commit c85c35ff built with OMLX_WITH_CUSTOM_KERNEL=1, MLX 0.32.2, one model loaded, one request at a time, 16-bit KV cache, MTP and thinking off, greedy decoding, context 8192 unless stated.

• —MLX allocator peak at 8K context: 39.45 GiB. Loading pages in 1.8–2.5 GiB of swap once; no swap growth was observed during generation at 8K, 32K or 64K.
• —The 129K figure is oMLX's own context-sizing benchmark (one output token); the largest chat-style request completed with a real answer was 64K + 256 tokens. Treat ~64K as the practical context on 48 GB and the 129K result as a prefill-capacity measurement.
• —Decode speed on this hardware is bound by kernel-launch overhead rather than memory bandwidth, so the ternary experts buy memory, not speed.

Quality

Three independent measurements, all with thinking disabled and greedy decoding.

KL divergence against a Q8_0 teacher

wikitext-2 test set, 40 windows of 2048 tokens (40,920 scored positions), llama.cpp's --kl-divergence convention. Teacher: lmstudio-community/Qwen3.8-Flash-Next-GGUF Q8_0 (perplexity 3.297). The MLX checkpoints were scored under MLX 0.32.2 on CUDA after losslessly re-packing the ternary groups to affine Q2, so all rows use the same stock kernels.

At this memory budget the total number of expert bits sets the KLD, not where they are placed: this checkpoint and the all-affine Q2 build are a statistical tie (per-position difference 0.012, standard error 0.003). The imatrix alone is worth 0.12 KLD over the weight-only ternary fit.

Knowledge and reasoning, 700 fixed cases

Stratified random subsets (seed 42) of MMLU (300 items, 5-shot), GSM8K (100), TruthfulQA MC1 (200) and CMMLU (100, 0-shot), served over the OpenAI-compatible API with 8192-token context, 128-token answers for multiple choice and 512 for GSM8K. Truncated, unparseable and errored answers count as wrong. The Q4KXL column was served by LM Studio (llama.cpp) on the identical cases.

Code generation (pass@1)

oMLX's bundled HumanEval and MBPP tasks, sandboxed execution, one attempt per problem.

Paired by problem: HumanEval both pass 150, only this 2, only Q4 7, neither 5; MBPP both 381, only this 9, only Q4 48, neither 62. Failures are mostly genuinely wrong code, not formatting.

Known issues and limitations

• —Chinese instruction following regressed. On CMMLU the model often writes an option-by-option analysis in Chinese instead of answering with the requested letter; 24 of the 100 CMMLU cases hit the 128-token cap that way (the previous ternary build did this 6 times). The analyses are usually correct, so it is a compliance loss rather than a knowledge loss, but it costs real points on strict-format Chinese tasks. The likely cause is the English/code-centric importance matrix now weighting the dense projections; a bake with the imatrix restricted to the experts is the planned bisect.
• —The KL divergence was measured on English text only. Chinese, other languages and code were not scored that way.
• —Vision inputs were not tested. The vision tower is present (its Linear layers at Q8), but no image prompt has been validated against the source model.
• —Thinking mode was disabled for every measurement here; tool calling and multi-request batching were not exercised.
• —Loading on a 48 GB machine briefly swaps 1.8–2.5 GiB. Keep other memory users closed, keep the memory guard on, and start with an 8K context before growing it.
• —The multi-token-prediction head is not included; speculative decoding gave little on this hardware and cost several GiB.

Running it

You need Apple Silicon, macOS 15 or newer, full Xcode (the Metal toolchain, not just Command Line Tools), uv and the Hugging Face CLI. Build the fork in its own checkout and environment; it is not a drop-in for an existing oMLX install.

git clone https://github.com/fuutott/omlx.git omlx-qwen48
cd omlx-qwen48
uv venv --python 3.12 .venv
OMLX_WITH_CUSTOM_KERNEL=1 uv pip install -e .
uv run python -c "from omlx.custom_kernels import native_kernel_status; print(native_kernel_status())"
# the 'bonsai' entry must report available: True

Download the checkpoint outside the Git checkout, verify it, and serve it:

hf download fuutott/Qwen3.8-Flash-Next-MLX-t5-imatrix-q3down-ple8 \
  --local-dir /path/to/checkpoints/Qwen3.8-Flash-Next-MLX-t5-imatrix-q3down-ple8

uv run python tools/quantize_qwen4_flash_next_t5.py \
  --verify-only /path/to/checkpoints/Qwen3.8-Flash-Next-MLX-t5-imatrix-q3down-ple8

uv run omlx serve --model-dir /path/to/checkpoints --memory-guard safe --memory-guard-gb 48

Put the checkpoint on a fast internal SSD: the 53.6 GiB PLE table is read through mmap on every token. In the oMLX admin UI confirm the model is detected as qwen4_exp and that Qwen4 PLE SSD offload is enabled (it is forced automatically at a 48 GiB ceiling). Keep MTP off, load one model at a time, and use context 8192 for the first run. The checkpoint validates with the converter's --verify-only (schema, packing, layout and the SSD-offload representation).

Reproducing the bake

The full recipe, environment lock and converter live in the fork's README under How to bake the cake. This checkpoint was produced on Windows with CUDA from the pinned source revision and the Unsloth importance matrix (imatrix_unsloth.gguf_file, SHA-256 a5863123db1ca458727e738955bef7bfc199520aa2bee3a30142a1aff9254154) with:

tools/quantize_qwen4_flash_next_t5.py --model <Qwen3.8-Flash-Next @ de4b8e4d> --output <dir>
  --expert-format t5 --t5-fitter prefix --expert-down-bits 3
  --imatrix <imatrix_unsloth.gguf_file> --imatrix-strict --imatrix-scope safe --allow-experimental-imatrix
  --clip-search --ple-bits 8 --vision-bits 8 --device cuda:0 --chunk-rows 4096

If you want the same size class without the fork's kernels, --expert-format affine produces the all-affine Q2 variant from the KLD table, which loads on stock oMLX.

Files

• —model-*.safetensors (131 shards) and model.safetensors.index.json
• —config.json, generation_config.json, tokenizer and preprocessor files, chat_template.jinja from the source model; config.json carries the omlx_t5 loader marker and declares no MTP layers
• —omlx_conversion.json: recipe, source identity, imatrix statistics and the per-tensor precision record
• —omlx_conversion_manifest.json: the converter's identity manifest (source, code and environment hashes, chunking) whose fingerprint is stamped into every shard's metadata
• —SHA256SUMS.txt: hashes of every file in this repository; check with shasum -a 256 -c SHA256SUMS.txt
• —LICENSE: the source model's Apache-2.0 license

Source: Qwen/Qwen3.8-Flash-Next at revision de4b8e4d43b917e7706784d8bb445c9af86a3540. Converter: fork commit c85c35ffbf24f83a085538266f3942eb68c43c1a.

License and credits

Apache-2.0, inherited from the source model; the base model's card and license conditions apply to these weights. Thanks to the Qwen team for the model, to oMLX for the runtime, the Bonsai ternary format and the Qwen4 PLE SSD offload this build depends on, to Unsloth for the importance matrix, to LM Studio for the Q8_0 teacher GGUF, to llama.cpp for the KL-divergence tooling, and to Tencent's AngelSlim work for the weighted ternary fitting idea.