qwen3.8-flash-next

Qwen3.8-Flash-Next HFQ — hipfire quants (MQ4 default, MQ3, MQ2)

Hipfire quantized builds of Qwen/Qwen3.8-Flash-Next for supported AMD routes. qwen3.8:flash is MQ4, the default. MQ3-XTS and the two MQ2 roots are smaller alternatives for tighter storage; each tier carries its own quality tradeoff, and no tier's numbers stand in for another's. Every base file needs one shared PLE file at serve time. Sizes are download sizes, not measured memory use. GB = 10⁹ bytes, GiB = 2³⁰ bytes.

Tiers and downloads

Totals are arithmetic file sums, not minimum RAM/VRAM.

Quickstart

hipfire pull qwen3.8:flash --ple auto
hipfire run qwen3.8:flash "…"

For another tier, use its tag in both commands: qwen3.8:flash-next-mq3-xts, qwen3.8:flash-next-mq2-xts, or qwen3.8:flash-next-mq2-xts-4500.

The PLE file holds lookup rows the base file points to; one of them is required at serve time. Q4 is the smaller download. Q8 is exact to the original Q8 rows, not to the BF16 source. --ple auto respects the configured PLE format; otherwise it selects q4 for gfx1151 and q8 elsewhere, combining the choices on mixed-GPU hosts. Recommended sampling: temperature 1.0, top_p 0.95, top_k 20.

Quality

KLD below is a top-K-screen distribution distance on the stated corpus and route. Lower means closer to that screen's reference — never task accuracy, never pooled across references, never a cross-tier ranking.

MQ4-XTS

Diagnostic pins with the trunk overlay OFF and IU4 OFF (build 7d4000765, q4 PLE, gfx1151, Top-256 + residual, BF16-source teacher):

These pins transfer by equivalent weights for the overlay-OFF diagnostic path only. No overlay-ON (gfx1151 default serving path) KLD is stated here. The older ≈ 0.066 figure (build fe77c0837, different route) is a different measurement and is not the current default claim.

The default base is a verified split of the pinned monolith: an independent 1,197-tensor walk is byte-equal, with only the 128 PLE shards moved out. (NLL/PPL/top-1 and kldseq SHAs for both pins live in kld/VERIFY.md, which covers MQ4 only.)

MQ3-XTS (R9700 / gfx1201, qwen4_kld build ca0aee9c9, q4)

BF16-reference rows and FP8-reference assistant rows are separate screens. The assistant corpus is a mixed assistant-proxy/HumanEval+OASST blend, not confound-fixed; its FP8 reference is not BF16. Paired predecessor→current deltas on matched chunks (WT2 +0.001317, code24 −0.002377, assistant decode −0.022933, cold prefill −0.054648) all pass the +0.010 one-sided point screen. Cold-prefill improvement is not statistically significant (CI crosses zero).

MQ3-XTS on 7900 XTX (diagnostic, separate scope)

Same baked bytes + q4, but a different build, arch, and forced offload env (HIPFIRE_OFFLOAD_EXEC=pcie + HIPFIRE_QWEN4_D2_DEFAULT=0, non-production). The WT2 subset differs from the R9700 full-32 run, so the two WT2 means are not directly comparable.

The shipped ca0aee9c9 binary refuses MQ3 on XTX (true for that build); these rows come from the later W3-slab build. Neither erases the other, and neither is an all-GPU claim.

MQ2 roots (7900 XTX, build ea2abd11…, q4, published baked roots)

Pre-land emulation (Table A, weight-only, FP8 basis — not a card measurement) comes first; on-card measurements (Table B) follow. The two tables are never pooled or ranked against MQ4/MQ3.

Six of eight rows were measured on the published baked roots. The two unmeasured rows are labelled "Not measured (identity + G2 pass)": runtime-consistency evidence (32 identical greedy ids, lane parity at k = 1, 2, 4, 8 with zero mismatches, 16/16 negative controls), not a KLD. Full paired deltas, CIs, and the 6.23% explicit exception are in the appendix. No speed, throughput, or memory claim.

Recipes

• MQ4-XTS: 6-bit trunk and 6-bit language head; Q8 embedding, MTP attention, and PLE n-gram rows; BF16 hyper-connection, router, shared expert, projections, and norms. The 108 4-bit trunk projections are not baked into the base — a per-arch sidecar applies them on gfx1151 by default and stays off on R9700 unless forced. A separate expert-hotset sidecar sets the R9700 default placement.
• MQ3-XTS: 16,839 cold experts at rotated group-128 W3 (16,679 re-solved
  • 160 RTN-fallback unchanged) plus 7,737 hot experts kept at MQ4, of 24,576 trunk experts. The 108 input projections ship baked as 4-bit QT44.
• MQ2 roots: 6,306 (or 4,500) hot experts kept at MQ4 plus the rest re-solved as E8P 2-bit lattice cold (one FP16 scale per 128 weights) on pooled d2 Hessians totaling 523,469 tokens. The 108 input projections ship baked as 4-bit QT44, like MQ3 and unlike MQ4.

Codec parameters, solver settings, and calibration hashes for each tier are in the technical appendix. The MQ4 calibration details there describe the MQ4 monolith only and are never copied onto MQ3/MQ2.

Compatibility

Build minima on the current 0.4.3 beta line: split MQ4 a6f241ce2, MQ3 ca0aee9c9, both MQ2 roots ea2abd11…. The ca0aee9c9 build loads MQ3 on R9700 but refuses it on 7900 XTX at admission (build-gated route, one file). R9700 KLD rows ran the unset-selector env as a quality screen, not a serving-route claim; production decode performance lives in separate w3perf-* results, never pooled with KLD.

Experimental: MQ3-XTS coder (k243)

Experimental, unregistered. A pruned code-mix quant for single-device work on the current beta. Not a registered tag — no hipfire pull tag and no alias/default change (qwen3.8:flash stays MQ4).

• Base qwen3.8-flash-next-base.mq3-xts-coder: 29.11 GB / 27.11 GiB (58.23 GB / 54.23 GiB with shared q4 PLE; 83.51 GB / 77.78 GiB with q8). Needs one shared PLE file at serve time, like every other tier.
• 243 kept experts per layer (11,664 W3 of 24,576 trunk experts), code-mix GPTQ re-solve — not finetuning.
• Held-out loss and partial benchmarks only (first-shard + explicit partial HumanEval; MBPP not completed; full-suite scores not completed). See mq3-xts-coder/README.md for the numbers, method, and exact non-claims.
• Requires the current 0.4.3 beta runtime on a single device; older builds are not compatibility-tested. Full details, usage, and hashes live in mq3-xts-coder/README.md.

<details><summary>Verification appendix (exact bytes, hashes, paired details, history)</summary>

Exact identities

MQ4 tags qwen3.8:flash-next, qwen3.8:flash-next-mq4, qwen3.8:flash-next-gptq3, and qwen3.8:flash-next-mq4-xts all point at the same eafaff94… base; qwen3.8:flash is an alias of qwen3.8:flash-next.

References: WT2 0314efabfbe3fd2c81c9a0530eac1fb725ca97c5238741c0ec16dcd480a95595, code24 0d6c6a8cf00062cfeafc4be206a44b3cb68622f9174fe852ff1d2bea99b1f28b, assistant FP8 00c9a24b558605839a9756d23ca12ba95b70480776b92897e2261c5845d249bd. kld/VERIFY.md and provenance.json cover MQ4 only — never cited as MQ3/MQ2 reproducers.

MQ2 paired details (XTX, evaluator 1df38c8d…, 20000-resample paired CIs)

Differences are XTX minus the corresponding pre-land reference, paired on the same chunks. The release rule accepts a non-significant increase or a significant increase within 6% of the reference mean; it uses the mean increase, not the upper confidence bound.

This 6% rule applies to this MQ2 release only. The exact-row exception does not change it. Evidence packet SHA-256: 190edcea411793fa9c881f1191cda9ae79b480a0b5ae394e5472bbfb2c4c3c71.

MQ3 methods

Paired bootstrap: 20,000 draws, seed 0, PCG64, chunk/sequence clusters, row-weighted, percentile 2.5/97.5. The d2 pre-ZAB base plus a runtime QT44 overlay scored identically to the baked root (overlay-equivalence; the raw pre-ZAB bytes were never scored alone). Full-corpus assistant rows are primary; excluding outlier sequence 63 gives decode 0.074163 → 0.055682 and prefill 0.070157 → 0.065096. XTX companion figures: NLL/PPL/top-1 WT2 1.826613 / 6.2128 / 0.8526, code24 1.590903 / 4.9082 / 0.8985.

Tier technical parameters

MQ4 (monolith provenance.json only): 240 MQ6G256V2 trunk/GDN projections + MQ6 head; symmetric MQ4 gate-up G256 / down G128 GPTQ experts (H2 = 1.0× mean diag, descending-diag permutation, 128-col blocks, ≤3 scale refits; no AWQ fold); calibration.u32 e00111ee…, 262,144 tokens in 128×2048 blocks, disjoint from WT2; 553 of 25,088 experts keep RTN. MQ3: gate-up MQ3I4G128R256 (qt131) + down MQ3I4G128R128 (qt132), group-128, symmetric 3-bit codes / f16 scales, rotation 256/128, grid ±1,3,5,7, scale rule mqi4-canonical-v1-rne-tree128; chat-aware GPTQ re-solve on d2 chat-mix Hessians (all 48 layers, 24,576 records verified, solve batch 128, λ = 1.0); 160 RTN-fallback unchanged. MQ2: E8P cold (2-bit E8-lattice codebook, one FP16 scale per 128 weights), solver vq_d2.py, pooled Hessians r1.train 237,568 + chat.p1 262,144 + cur.p1 23,757 = 523,469 tokens; hot MQ4 tiers unchanged. The 108-QT44 bake is verified for both MQ2 roots (108 replacements, 50,145 unchanged tensors). Pre-bake d2 bases are superseded by the published baked roots.

History

Current files for new pulls: the four split bases, the two shared PLE files, and the MQ4 sidecars above. Older monoliths (qwen3.8-flash-next.mq4-xts canonical name from hipfire 0.4.1.1, legacy …-gptq3.mq4 alias, RTN-asymmetric ….mq4, mq6q8-pleq8) remain hosted for reproducibility; use current tags for new pulls. Older d2 bytes stay recoverable in history but are never measurements of the new hashes. Full verbatim history: README.md @ 0e5e9fd9.

License

Source: Qwen/Qwen3.8-Flash-Next at de4b8e4d43b917e7706784d8bb445c9af86a3540. Quantization and runtime by hipfire (Kaden Schutt / warpfront). Model weights remain under the Qwen Community License 1.0 — not Apache/MIT, which cover the hipfire container and runtime work only.

</details>