Promoted active in llama-swap.yaml · aliases: qwen3-thinking-80b
| Run | Score | tok/step | R5 | Log |
|---|---|---|---|---|
| No results-*.log found for this id/alias. | ||||
The 3x repeat battery results confirm this is the new top-tier model on the box, surpassing [[coder-agentic-regrilled]] on almost every domain-specific gate while maintaining comparable throughput.
| suite | Score | Notes |
|---|---|---|
| R1 (Coding) | 22/24 | Extremely consistent; matches coder-agentic's ceiling. |
| R2 (Coding) | 14/15 | One minor regression from rep1. |
| R3 (Coding) | 15/15 | Perfect performance. |
| R4 (Coding) | 14/15 | Consistent with R1/R2. |
| R5 (Coding) | 15/18 | 0/18 looped; clean execution despite 3 redundancies. |
| KLayout+ref | 20/24 (83.3%) | Beats coder-agentic's 81.3% (13/16). |
| Office (both) | 16/27 (59.2%) | Hits the vision floor (6/9) for text-only models. |
| Realcase | 11/21 (52.4%) | Significant win over coder-agentic's ~30-40%. |
Aggregate Coding (R1-R5): 80/87 (~92%) vs coder-agentic's ~93.5%.
Throughput: ~73 tok/s (matches/beats coder-agentic's ~70 tok/s).
Looping: 0/18 total loops across all reps.
[[coder-agentic-regrilled]] is slightly more "stable" on R1/R2 coding, but qwen3next-thinking is the superior tool for KLayout and real-world industrial tasks (Realcase). It offers a more "intelligent" reasoning path with zero loop pathology.[[llama4-scout-17b-16e-tq1]] is significantly weaker on coding and realcase, though MoE robustness at extreme quant is confirmed.This model is highly recommended for agentic workflows where KLayout/Excel/Real-case accuracy is prioritized over absolute coding speed. It is the current "thinking" alternative to coder-agentic.
Why: High correctness across all suites, beats the previous driver on the most critical (real-use) gate, and maintains the established throughput/context budget.
How to apply: Add to llama-swap.yaml. Use as the preferred driver for KLayout and Realcase tasks.
"qwen3next-thinking":
aliases: [qwen3-thinking-80b]
# Qwen3-Next-80B-A3B-Thinking, UD-IQ2_XXS GGUF (unsloth), llama.cpp.
# Direct comparison to coder-agentic (43/46 coding, klayout+ref 81.3%, realcase ~30-40%).
# BEATS coder-agentic on KLayout (83.3%) and Realcase (52.4%).
# 262144 ctx, ~73 tok/s.
# REASONING BUDGET: a --reasoning-budget 4096 cap was tried 2026-08-16 to
# quiet console output, then REVERTED the same day: the battery's own
# results.jsonl shows realcase alone used 7693/9248/11864 total out_tok
# across its 3 reps (the suite that most drove the PROMOTED verdict,
# 52.4% vs coder-agentic's ~30-40%), and 25/69 logged tasks exceeded 4096
# total tokens. A 4096 reasoning cap would very plausibly have truncated
# thinking mid-task on exactly the suite the promotion rests on. Left
# UNRESTRICTED (the default, matches what every recorded score measured)
# -- console noise is a real cost but re-measuring after a cap requires
# re-running realcase+klayout to know the actual quality impact, not
# picking a number and hoping. See qwen3next-thinking-verdict.md.
cmd: |
/usr/bin/env
CUDA_DEVICE_ORDER=PCI_BUS_ID
${llama_bin}
# STORAGE /mnt/models (USB SSD, 33% used) — MOVED off ssk500 2026-08-25 to
# relieve that drive at 91%. Both externals measure ~390 MB/s, so tier by
# CAPACITY not speed ([[storage-tiers]], [[mnt-models-usb-ssd]]): this move
# costs nothing in load time, unlike an NVMe demotion (~+47 s).
-m /mnt/models/gguf/Qwen3-Next-80B-A3B-Thinking-UD-IQ2_XXS.gguf
--alias qwen3next-thinking
--jinja -ngl 99 -c 262144 -fa on
--tensor-split 48,52
-b 2048 -ub 1024
--temp 0.2 --top-p 0.9 --top-k 40 --repeat-penalty 1.05
--cache-type-k q4_0 --cache-type-v q4_0
--host 127.0.0.1 --port 9145 --parallel 1
proxy: http://127.0.0.1:9145