Qwen3.8-Flash-Next for mlx-serve, iQ-MLX 3.3 bpw (64 GB Macs)

mlx-serve pack of Qwen/Qwen3.8-Flash-Next, the Qwen4 preview architecture (model_type: qwen4_exp), quantized to 3.3 bits per weight with imatrix calibration so it fits a 64 GB Mac with room for 64k of context. 52 GB resident. Includes the MTP head and the vision tower (image and video input).

Download MLXServe.com

sudo sysctl iogpu.wired_limit_mb=58000     # 64 GB Macs, once per boot; see below
mlx-serve --model ddalcu/Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-3.3bpw --serve --ctx-size 65536 --skip-mem-preflight

(launching mlx-serve via UI is also fine, make sure to enable --skip-mem-preflight via settings)

Pick the pack by your Mac's memory:

Download includes the n-gram table, which is mmapped and not resident (see below).

Quality

Measured against the bf16 checkpoint on 1200 held-out next-token positions (agent, code, prose, math; documents the calibration never saw). Top-1 is how often the pack's greedy token equals bf16's; mass is the bf16 probability of the pack's pick, with bf16's own pick as the ceiling.

Same positions, same server, cold prefills. The 3.3 bpw pack keeps 99% of the 4-8 pack's probability mass at 21 GB less.

Variants tried and rejected: more expert bytes at the cost of a 6-bit spine (worse), forcing the down projections to 3-bit (a wash), 3.6x more calibration data at 44 GB (no change in measured error; bytes moved the needle, calibration volume did not).

Memory on a 64 GB Mac

macOS caps what the GPU may wire at about 75% of RAM (48 GB on a 64 GB Mac), which is not enough. Raise it once per boot:

sudo sysctl iogpu.wired_limit_mb=58000

With that, 52 GB of weights plus a 64k KV cache (1.6 GB) plus the prefill working set fit. Measured on this pack: 52.2 GB active after load, 55.6 GB peak during a 36k-token prefill. Keep --ctx-size at 65536; --kv-quant 8 halves the cache if you need more headroom. The 32 GB n-gram table is not in that number (next section).

What is different about this model

This is not a Qwen3.5-style pack. Three things around the usual GDN + MoE trunk:

• Gated residual streams. The residual is 4 streams wide (4 x 2560). Every block reads a sigmoid-mixed average of the normalized streams and writes back through per-stream scalar gates. The final mixer replaces the usual final norm.
• N-gram embedding (51B parameters). A second embedding table indexed by hashed bigrams and trigrams of the token ids: 16 heads, each a prime-sized bucket space of ~20M rows, 160 dims per row, injected once before layer 1. It is a lookup, no compute, which is why Qwen quotes the model as 125B: the full checkpoint is 125B trunk + 51B n-gram + 4B MTP = 180B (360 GB bf16).
• Qwen Sparse Attention. Past 2048 tokens each attention layer only reads the 512 most relevant 4-token blocks per query (picked by a small indexer), plus the query's own partial block. Attention cost stays flat with context. Native 262k context.

How this pack stores the n-gram table

The 51B table is NOT in the safetensors shards. It is one merged 4-bit table in ngram_table.bin (32.0 GB, safetensors format, .bin so nothing mlx-loads it). mlx-serve mmaps the file and, per token, dequantizes the 16 rows it needs on the CPU (16 x 80 bytes) and uploads only the resulting 2560-vector. The table never becomes resident: its cost is page cache, which the OS evicts as needed. On a 64 GB Mac that page cache competes with everything else, so the first long prompt after boot pays random SSD reads (a few seconds per 8k tokens on a cold cache); warm is free.

Widths

Every quantized weight with an imatrix entry uses an activation-weighted (scale, bias) search instead of min-max, byte-compatible with mx.quantize's affine layout, so the engine reads it like any other affine pack. Every (1 + w) RMSNorm has the +1 folded into the stored weight; depthwise convs are transposed to MLX's [C, K, 1]; experts.gate_up_proj is split into switch_mlp.gate_proj / up_proj. The vision tower ships dense bf16 in model-vision.safetensors (~0.9 GB).

Serving notes

• Speed. M4 Max, llmprobe --bench-only --rungs 2k: 52.6 tok/s decode, 762 tok/s prefill, 318 ms TTFT; the mixed-4-8bit pack measured 55.5 / 754 / 278 ms in the same run. 3-bit experts ride MLX's stock gather kernel, 2-bit ones mlx-serve's fused ones.
• MTP. The checkpoint's own 1-layer speculative head is in the pack (--mtp or per-request "enable_mtp": true). Default-off; not re-measured on this pack.
• Thinking is on by default ("enable_thinking": false turns it off). Tools use Qwen3.8's XML call format; mlx-serve parses and schema-coerces it.
• Images and video go through the Qwen3-VL-style tower (model.visual.*, dense bf16). MTP is declined on image turns (serial decode).

Calibration and conversion

All in the mlx-serve repo, tests/:

1. qwen38_flash_next_imatrix_collect.py collects per-input-channel activation statistics on the exact bf16 checkpoint (482k tokens of agent traffic, code, prose and math, rendered with the model's own chat template) by streaming one decoder layer at a time through HF transformers, so the 360 GB never has to be resident. The same script writes the held-out bf16 logits the quality table is scored on.
2. qwen38_flash_next_iq_allocate.py measures the imatrix-weighted error of every layer's gate-up and down banks at 2/3/4 bits and spends a 47 GB budget where it buys the most.
3. convert_qwen38_flash_next.py --imatrix --alloc writes the pack.
4. qwen38_flash_next_score.py scores any served pack against the reference.