Ecmcfee/Swift1.5-Qwen3.8-Flash-Next-q4-MTPLX

Swift 1.5 Qwen3.8-Flash-Next — MTPLX "Optimized Speed" quant (q4, MTP + n-gram)

This is a quantized MLX pack of ukisai/Swift1.5-Qwen3.8-Flash-Next — UkisAI's reasoning-efficient finetune of Qwen3.8-Flash-Next — produced by applying the MTPLX "Optimized Speed" quantization recipe from Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed to the Swift-1.5 weights.

It was forged locally with mtplx forge (MTPLX 2.12.2, qwen4-exp lane) from source revision 0bd4fe22431372cdad1979267d3ab45aa7e6150a, full-load verified before stamping, and runs with MTP draft-acceptance active and the n-gram table offloaded to SSD — the same target class as the original pack: M5 Ultra 96 GB or M5 Max 128 GB Macs with fast NVMe storage.

What this pack is

Component	File	Precision	Notes
Body (48 layers: 36 GDN linear-attn + 12 full-attn, 512-expert MoE, hidden 2560)	model-0000[1-9]-of-00019.safetensors	4-bit affine, group 64 (~4.25 bpw effective)	per-tensor map in config.json
8-bit promotions	(inside body shards)	8-bit affine, group 64	embed_tokens, lm_head, linear_attn.out_proj, mlp.gate (router), shared_expert_gate, shared_expert.{gate,up,down}_proj, indexer.index_qk_proj
Kept BF16	(inside body shards)	bf16	hyper-connection input_mix_weight_down/up, block_inject_weight, ple.key_proj/value_proj, linear_attn.conv1d
MTP head (depth 3)	mtp.safetensors (1.6 GB)	4-bit, group 64	fast-path layout pinned to 4-bit; MTP chain probe: ~110 tok/s vs ~52 tok/s AR (2.1×)
N-gram PLE table	ngram-table.safetensors (29.8 GiB)	4-bit, group 32	streamed from SSD, not RAM-resident — serves on 96/128 GB machines
Vision tower	model-vision.safetensors (0.9 GB)	bf16	333 tensors; multimodal (image-text-to-text) preserved

Note vs the original Youssofal pack: this build runs qsa_8bit: false (full-attention self_attn.{q,k,v,o}_proj stay 4-bit) and group size 64 on the body; the 8-bit promotions listed above match the recipe.

Requirements & serving

• Hardware: designed for M5 Ultra 96 GB / M5 Max 128 GB (resident footprint ≈ 83 GiB incl. MTP, verified on 128 GB; the 29.8 GiB n-gram table streams from SSD). Do not expect it to fit 64 GB machines.
• Runtime: MTPLX / MLX qwen4-exp lane (MTPLX ≥ 2.12.2). This is not a transformers-compatible checkpoint; the config.json quantization map is MLX layout.
• Serving notes: keep the n-gram table on NVMe; recommended sampler temperature 1.0, top_p 0.95, top_k 20; MTP depth 3 with mtp_position_mode: cache (calibration kept the family-default contract).

Provenance & verification

• Source: ukisai/Swift1.5-Qwen3.8-Flash-Next (BF16 native, MTP + n-gram + vision shards in-repo), sha 0bd4fe22….
• Forged locally 2026-10-14 with mtplx forge build (speed recipe); full-load verified (mtplx_runtime.json → verification: verified; forge verify rows: AR 52.3 tok/s, MTP d3 110.5 / 104.8 tok/s).
• Unofficial community conversion — no affiliation with UkisAI, Qwen, or Youssofal.

License

The underlying model is distributed by UkisAI under the Swift Open License v1.0 (includes enterprise clauses) plus Qwen License terms — quantization changes nothing about that: check both licenses (and the ukisai repo's LICENSE) before any commercial use. Personal/local use is fine.