Qwen3.8-Flash-Next-heretic-2 — GGUF (static quants)

🧬 Source model: trohrbaugh/Qwen3.8-Flash-Next-heretic-2 · base: Qwen/Qwen3.8-Flash-Next
📦 imatrix (i1) quants: spiritfather/Qwen3.8-Flash-Next-heretic-2-i1-GGUF · MTP head: mtp/ · vision: mmproj/
🛠️ Runs on llama.cpp (master); MTP needs PR #27836 · decensored with Heretic
📊 Scored on CaliperBench
💬 Discord

[!Note] All credit for the model goes to trohrbaugh, who produced Qwen3.8-Flash-Next-heretic-2 with a custom fork of Heretic (v1.3.0+custom, per-layer direction) — author-reported 0/100 refusals (base 99/100), KL divergence 0.0818, the lowest-KL heretic of Qwen3.8-Flash-Next at time of writing. The base model is Qwen/Qwen3.8-Flash-Next (177B total / ~3B active MoE, arch qwen4exp) by the Qwen team. This repository only re-packages those weights as GGUF.

[!Tip] Benchmarked on CaliperBench — a creative-writing benchmark scoring prose craft, roleplay and willingness rather than general intelligence. See this model's scores: caliperbench.com/m/qwen3.8-flash-next-heretic-2/.

[!Important] These files ship the model's own MTP/NextN speculative-decoding head (mtp/, plus a pre-grafted pair per quant) — the first Flash-Next quants to do so. The trunk files load on any current llama.cpp; the head needs PR #27836. See MTP.

Provided quants

PPL/KLD: llama-perplexity on wikitext-2-raw test, ctx 512, 120 chunks, reference logits from this repo's Q8_0 (the bf16 trunk is 250 GiB and does not fit a 256 GB Mac; Q8_0 is within noise of it). Lower KLD is better; AesSedai's stock-model table uses the same method against bf16 and is the closest comparison.

Loading (any current llama.cpp)

The trunk files load on stock llama.cpp — qwen4exp merged upstream in #27742 (b10685+). Vision needs --mmproj mmproj/mmproj-Qwen3.8-Flash-Next-heretic-2-Q8_0.gguf (projector is byte-identical to base; heretic only edits the LM).

llama-server -m Q4_K_M/Qwen3.8-Flash-Next-heretic-2-Q4_K_M-00001-of-00005.gguf -ngl 99 -fa on -c 65536 -ctk q8_0 -ctv q8_0 --jinja

Memory: only the trunk lives on the GPU (the "On GPU" column). The 51B-row n-gram table per_layer_token_embd (50.7 GiB at Q8_0) is a lookup table; with --tensor-read-lazy auto (default on builds with #27837) it stays on disk and only touched rows page in — measured ~47 GB less resident memory, zero throughput cost. Context is cheap: ~21 KiB/token with q8_0 KV (full native 262K ≈ 5.5 GiB).

Thinking: on by default (reasoning_effort xhigh|medium|low). Turn off per request with chat_template_kwargs: {"enable_thinking": false} or serve with --reasoning off --chat-template-kwargs '{"enable_thinking": false}'.

MTP speculative decoding (the model's own draft head)

Qwen ships a 4B NextN/MTP head trained jointly with the model; most Flash-Next GGUFs drop it. Here it is kept:

• mtp/mtp-Qwen3.8-Flash-Next-heretic-2-Q8_0.gguf — the head alone (34 tensors, 3.85 GiB), exported with convert_hf_to_gguf.py --mtp.
• <QUANT>/MTP/ — a pre-grafted shard pair (rewritten shard 1 + a head shard) so you can add MTP to a quant you already downloaded without re-downloading the trunk.

Requires llama.cpp PR #27836 (qwen4exp : add NextN/MTP draft head) — stock master does not load the head for this arch (it ignores nextn_predict_layers for qwen4exp). Files follow that PR's tensor layout (blk.48.*, nextn.hc_head_*, +1-shifted norms) as of commit 1d8de7c; if the layout changes before merge these two small artifacts will be re-exported. The head must be in the model file — -md mtp-….gguf does not work on that PR (it treats the sidecar as a full model). A build without the PR simply ignores the head tensors.

Steps (example Q4_K_M):

git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
git fetch origin pull/27836/head:pr27836 && git checkout pr27836
cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j --target llama-server

# next to the trunk shards, add the two MTP files, then clone/rename the middle shards to the -MTP- pattern (cp -c = APFS clone; use ln on Linux)
cd Q4_K_M/ && cp MTP/* .
N=$(ls Qwen3.8-Flash-Next-heretic-2-Q4_K_M-MTP-00001-of-*.gguf | sed -E 's/.*-of-0*([0-9]+)\.gguf/\1/')
for i in $(seq 2 $((N-1))); do cp -c Qwen3.8-Flash-Next-heretic-2-Q4_K_M-$(printf %05d $i)-of-$(printf %05d $((N-1))).gguf Qwen3.8-Flash-Next-heretic-2-Q4_K_M-MTP-$(printf %05d $i)-of-$(printf %05d $N).gguf; done

llama-server -m Qwen3.8-Flash-Next-heretic-2-Q4_K_M-MTP-00001-of-$(printf %05d $N).gguf -ngl 99 -fa on \
    --spec-type draft-mtp --spec-draft-n-max 5 --spec-draft-p-min 0.75      # 6-7 for Q4-class quants; see the sweep below

Check the log for creating MTP draft context and draft acceptance = … lines.

MTP report

Settings sweep on the stock Qwen3.8-Flash-Next (UD-Q4_K_XL + Q8_0 head, Apple M3 Ultra 256 GB, Metal, ~975-token prompt, 300 generated tokens, temp 0, single slot):

Sweep on this repo's Q8_0 (same box, same prompts; plain decode 36.1 code / 36.0 prose):

(p-min 1.0 never drafts: flat 33 t/s, i.e. plain minus overhead.)

Takeaways: keep --spec-draft-p-min set (0.75–0.9); without it, depth beyond 3 loses to plain decode on prose. On Metal the sweet spot is n-max 5–6 for Q8_0 (+70% code, +18% prose) and 6–7 for Q4 quants; one step past that falls off a cliff (verify-batch threshold in the Metal MoE path — may differ per backend). Reported elsewhere on this PR: HIP +17% at n-max 2; CUDA works; Vulkan is a net loss (rejected drafts replay the recurrent state). Greedy output is close to but not bit-identical with plain decode on any backend.

On these heretic-2 files: measured on this repo's Q4_K_M (Apple M3 Ultra 256 GB, Metal, PR #27836 build, ~975-token prompt, 300 generated tokens, temp 0, single slot, n-max 7 / p-min 0.75): code: 28.6 → 49.9 t/s (+75%), draft acceptance 86% (227/263); prose: 26.7 → 35.4 t/s (+33%), draft acceptance 80% (151/188). The head carries over to the heretic intact — acceptance matches the stock model's. The trunk was also confirmed to load and generate on a stock upstream-master build (90c26fc, no PRs), where the head files are simply ignored.

Quant recipe

llama-quantize from a bf16 trunk exported with --no-nextn. Only the routed-expert tensors (ffn_{gate,up,down}_exps) carry the nominal quant; everything else is pinned to Q8_0 (--tensor-type on attn_, ssm_, hc_, _shexp, ple_, per_layer_token_embd, plus --token-embedding-type q8_0 --output-tensor-type q8_0). This is the same shape AesSedai and unsloth use for this arch, and it matters: a plain llama-quantize Q4_K_M (which drops the 2-KV-head attention, the rank-320 hyper-connections, ssm_out, shared experts and embeddings to Q4–Q6) measured KLD 0.177 / PPL +9.1% vs Q8_0 on this model; the pinned recipe costs ~0.5 GiB more. ffn_down_exps has 640 columns and cannot take 256-block K/IQ quants, so llama.cpp upgrades it automatically (Q8_0/Q5_0 by layer in Q4_K_M, Q8_0 in Q6_K/Q5_K_M, IQ4_NL in IQ quants). Indexer projections stay BF16, gates/norms F32. Head quantized separately to Q8_0; grafted with jlkivey's graft-mtp-shard.py (byte-exact, verified).

Built with llama.cpp qwen4exp/lazy-tensor+mtp = master 90c26fc + PR #27836 + PR #27837, transformers 5.16.1.

License

Qwen Community License 1.0, inherited from the base model (not Apache-2.0: Model-as-a-Service use needs a separate license from Qwen).