Gemma 4 26B-A4B 98e v7-coder — science-augmented code prune

20.8B params · 98 experts (30 dropped) · ~4B active · code map + targeted_gpqa

A research checkpoint that takes Gemma-4-26B-A4B-it and drops ³⁰⁄₁₂₈ experts per layer using a code-targeted recipe on the rebuilt v7 competence maps (audited producers, 10 classes, multilingual category), on a [24,40] per-layer floor — plus a targeted_gpqa class at weight 1.5 that protects a science-specialist keep-set derived from GPQA-diamond pass-traces. Same router, attention, and norms as base, plus the mandatory shared-FFN α=1.2 upweight every coder variant carries.

A coder that kept all its science: it holds v6-coder’s top-tier code profile while pulling GPQA-diamond to 70.71% — +9.6pp over v6-coder and at parity with the unpruned 128e (67.17 on the same Q6_K run). For maximal coding throughput with science left at baseline, see the sibling v7-coderx.

Full model card & methodology: ManniX-ITA/gemma-4-A4B-98e-v7-coder-it on Hugging Face.

Other formats: - GGUF (29 tiers, imatrix, CD-* per-layer mixes + F16 + mmproj): ManniX-ITA/gemma-4-A4B-98e-v7-coder-it-GGUF - NVFP4A16 (native vLLM, ~13 GB): ManniX-ITA/gemma-4-A4B-98e-v7-coder-NVFP4A16

Scores (Q6_K, llama.cpp, greedy, same host)

Reference columns on the same Q6_K run: unpruned 128e GPQA 67.17 / AIME 73.33 / HE 97.56 / LCB-55 96.36; v6-coder GPQA 61.11 / HE 98.17 / LCB-55 92.73. v7-coder matches or beats 128e on GPQA, AIME and GSM8K while holding the cohort code profile. (Small benches — GPQA 198q, AIME 30q — carry run-to-run variance; read GPQA/AIME as “recovered the science gap”, not a robust win over the base.)

Quantizations — HE+ / MultiPL-E-100 score, size & answer length

Every K-quant and CD tier was scored on HumanEval+ (164) and MultiPL-E-100 (llama.cpp, greedy T=0), with per-problem completion length from token_stats. bpw is the true bits-per-weight (8 × bytes ÷ 19,877,953,946). ⭐ marks a recommended pick.

Recommended picks:

• Q6_K ⭐ (17.81 GB) — max quality — 93.29% HE+ (17.8 GB).
• qat-Q4_0 ⭐ (11.42 GB) — ties the top HE+ (93.29%) at 11.4 GB — QAT base.
• CD-qat-Q4_K_M ⭐ (10.83 GB) — recommended default — 92.68% HE+ at 10.8 GB.
• CD-Q2_K ⭐ (8.82 GB) — smallest tier still in the 90%+ band — 91.46% HE+ at 8.8 GB.

The K-quant and CD tiers hold HE+ in the 90–93% band with length essentially identical to Q6_K; the 2-bit Q2_K_L / IQ2_XS tiers are the cliff (HE+ into the 80s/70s, token p90 blows out). The K-quant CD tiers are the recommended low-bit path — CD-IQ* i-quant bodies are not offered (the pruned MoE degenerates on an IQ-family body, score → 0). Prefer Q4_K_M or higher for production.

Head-to-head by file size — v7-coder vs Qwen2.5-Coder-14B (iso-disk)

Pairing by tier name is misleading — this is a ~20.8B-total MoE and Qwen2.5-Coder-14B is a 14.7B dense model, so the same tier name lands at a different size. The fair comparison is iso-disk: at a given GB budget, which scores higher on HumanEval+? Same rig (RTX 3090, llama.cpp, greedy). Qwen GGUFs are bartowski’s (83–85% across the ladder). At every band the MoE runs lower bpw at the same disk and still scores higher.

Pull

ollama pull mannix/gemma4-98e-v7-coder                  # :latest = Q4_K_M
ollama pull mannix/gemma4-98e-v7-coder:CD-qat-Q4_K_M    # ⭐ recommended — 92.68% HE+, 10.8 GB
ollama pull mannix/gemma4-98e-v7-coder:Q6_K             # max quality — 93.29% HE+ (bench tier)
ollama pull mannix/gemma4-98e-v7-coder:CD-Q2_K          # smallest 90%+ — 91.46% HE+, 8.8 GB
ollama pull mannix/gemma4-98e-v7-coder:vision-Q4_K_M    # + SigLIP vision tower

Inherits Gemma 4’s thinking format — serve with the reasoning parser enabled (--reasoning-format deepseek --reasoning-budget 8192 on llama-server).

Derivative of Gemma 4 — Gemma Terms of Use.