189 Downloads Updated 5 hours ago
ollama run mannix/gemma4-98e-v7-coder:Q8_0
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
| GPQA-D | AIME | MATH-500 | GSM8K | ARC | IFEval | HE | HE+ | LCB-55 | LCB-100 | MultiPL-E |
|---|---|---|---|---|---|---|---|---|---|---|
| 70.71 | 76.67 | 92.00 | 93.00 | 94.80 | 95.00 | 98.78 | 92.68 | 96.36 | 97.00 | 88.67 |
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.)
Per-tier HE+ was not swept for the v7 cohort — the Q6_K table above is the reference; quality below ~Q3 / 3-bit degrades on the Gemma 4 MoE, so prefer Q4_K_M or higher for production. Full 29-tier list (incl. CD-* per-layer mixes) is on the GGUF repo.
| Tier | Size (GB) | bpw | Role |
|---|---|---|---|
| Q8_0 | 21.16 | 8.14 | near-lossless |
| Q6_K | 17.81 | 6.84 | max fidelity (bench tier) |
| Q5_K_M | 15.07 | 5.79 | high quality |
| Q4_K_M | 13.24 | 5.09 | :latest — recommended default |
| Q4_K_S | 12.21 | 4.69 | compact 4-bit |
| ⭐ IQ4_XS | 11.01 | 4.23 | safe sub-12 GB 4-bit |
| Q3_K_M | 10.51 | 4.04 | smallest comfortable |
| IQ3_XS | 9.22 | 3.54 | sub-10 GB |
| IQ2_M | 8.22 | 3.16 | sub-8.5 GB (degraded) |
| IQ2_S | 7.83 | 3.01 | smallest viable |
ollama pull mannix/gemma4-98e-v7-coder # :latest = Q4_K_M
ollama pull mannix/gemma4-98e-v7-coder:IQ4_XS # safe 4-bit, sub-12 GB
ollama pull mannix/gemma4-98e-v7-coder:Q6_K # max fidelity (bench tier)
ollama pull mannix/gemma4-98e-v7-coder:IQ2_S # sub-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.