Qwen3.8-27B not showing up for multiple reasons

[matusfaro]

Description

Hello,

I was wondering why Qwen3.8-27B wasn't showing up for me. Claude was able to find some reasons why it gets deprioritized in the current code, which I believe should not happen on a default run, it hides models I'd like to see:

1. Qwen/Qwen3.8-27B is tagged pipeline_tag image-text-to-text on HF (natively multimodal) and doesn't get fetched by default unless you specify include_vision=True

  # cli.py:408
  def _include_vision_candidates(profile: str) -> bool:
      return profile.lower() in {"vision", "any"}

Minor versions are collapsed into one family. qwen3.8-27b, qwen3.6-27b, qwen3.5-27b all normalize to qwen3-27b. Inside engine/ranker.py:837 it picks qwen3.6 over qwen3.8 for several reasons:

2. Benchmark data is frozen pre-release:

  Qwen/Qwen3.6-27B   quality=88.95  status=direct     selkey=93.95
  Qwen/Qwen3.8-27B   quality=47.80  status=estimated  selkey=47.80

3. Lineage table penalizes it since it's a new model. qwen3.8-27b falls through to the bare qwen3 pattern:

 # data/lineage.py
  (r"qwen3\.6", 7), (r"qwen3\.5", 6), ..., (r"qwen3", 5),

Steps to Reproduce

uvx whichllm@latest

Hardware Info

➜  ~ uvx whichllm@latest

╭────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────── Hardware Info ───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ GPU 0: Apple M2 Max — 96.0 GB shared (budget 94.0 GB) — BW: 400 GB/s                                                                                                                                                                                                                       │
│ CPU: Apple M2 Max — 12 cores                                                                                                                                                                                                                                                               │
│ RAM: 96.0 GB                                                                                                                                                                                                                                                                               │
│ Disk free: 33.6 GB                                                                                                                                                                                                                                                                         │
│ OS: darwin                                                                                                                                                                                                                                                                                 │
│ VRAM headroom: 2.0 GB reserved per GPU                                                                                                                                                                                                                                                     │
╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

                                            Recommended Models                                            
┏━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━┓
┃     ┃                                          ┃        ┃  Fit /   ┃              ┃            ┃       ┃
┃   # ┃ Model                                    ┃ Quant  ┃   VRAM   ┃        Speed ┃ Published  ┃ Score ┃
┡━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━┩
│   1 │ Qwen/Qwen3.6-27B                         │ Q5_K_M │ Full GPU │  8.9 tok/s ~ │ 2026-04-21 │  88.9 │
│     │ 27.8B                                    │        │ 21.2 GB  │              │            │       │
├─────┼──────────────────────────────────────────┼────────┼──────────┼──────────────┼────────────┼───────┤
│   2 │ google/gemma-4-31B-it                    │ Q4_K_M │ Full GPU │ 10.3 tok/s ~ │ 2026-03-11 │  85.1 │
│     │ 31.3B                                    │        │ 20.1 GB  │              │            │       │
├─────┼──────────────────────────────────────────┼────────┼──────────┼──────────────┼────────────┼───────┤
│   3 │ google/gemma-4-26B-A4B-it                │  Q8_0  │ Full GPU │ 36.6 tok/s ? │ 2026-03-11 │  85.0 │
│     │ 25.8B (3.8Ba)                            │        │ 27.0 GB  │              │            │       │
├─────┼──────────────────────────────────────────┼────────┼──────────┼──────────────┼────────────┼───────┤
│   4 │ Qwen/Qwen3-30B-A3B                       │  Q8_0  │ Full GPU │ 46.3 tok/s ? │ 2025-04-27 │  84.3 │
│     │ 30.5B (3.0Ba)                            │        │ 31.6 GB  │              │            │       │
├─────┼──────────────────────────────────────────┼────────┼──────────┼──────────────┼────────────┼───────┤
│   5 │ Qwen/Qwen3-Next-80B-A3B-Instruct         │ Q3_K_M │ Full GPU │ 59.0 tok/s ? │ 2025-09-09 │  79.0 │
│     │ 81.3B (3.0Ba)                            │        │ 34.5 GB  │              │            │       │
├─────┼──────────────────────────────────────────┼────────┼──────────┼──────────────┼────────────┼───────┤
│   6 │ openai/gpt-oss-20b                       │  Q8_0  │ Full GPU │ 38.6 tok/s ? │ 2025-08-04 │  78.7 │
│     │ 20.9B (3.6Ba)                            │        │ 22.0 GB  │              │            │       │
├─────┼──────────────────────────────────────────┼────────┼──────────┼──────────────┼────────────┼───────┤
│   7 │ zai-org/GLM-4.7-Flash                    │  Q6_K  │ Full GPU │ 16.8 tok/s ? │ 2026-01-19 │  76.3 │
│     │ 31.2B (12.0Ba)                           │        │ 26.2 GB  │              │            │       │
├─────┼──────────────────────────────────────────┼────────┼──────────┼──────────────┼────────────┼───────┤
│   8 │ Qwen/QwQ-32B                             │ Q4_K_M │ Full GPU │  9.8 tok/s ~ │ 2025-03-05 │  74.8 │
│     │ 32.8B                                    │        │ 21.0 GB  │              │            │       │
├─────┼──────────────────────────────────────────┼────────┼──────────┼──────────────┼────────────┼───────┤
│   9 │ deepseek-ai/DeepSeek-R1-Distill-Qwen-32B │ Q4_K_M │ Full GPU │  9.8 tok/s ~ │ 2025-01-20 │  73.8 │
│     │ 32.8B                                    │        │ 21.0 GB  │              │            │       │
├─────┼──────────────────────────────────────────┼────────┼──────────┼──────────────┼────────────┼───────┤
│  10 │ Qwen/Qwen3-14B                           │  Q6_K  │ Full GPU │ 13.7 tok/s ~ │ 2025-04-27 │  73.7 │
│     │ 14.8B                                    │        │ 13.5 GB  │              │            │       │
└─────┴──────────────────────────────────────────┴────────┴──────────┴──────────────┴────────────┴───────┘
  Speed:  ~ = estimated tok/s range,  ? = low-confidence/backend-sensitive tok/s
  Top pick confidence: High (direct benchmark, gap +3.9)
  Benchmark reference: 2026-05 curated snapshot; live AA / LiveBench / Aider merged when reachable.
  Speed caution: Low-confidence speed estimates in top ranks: #3

Python Version

No response

Operating System

macos 27

whichllm Version

0.5.16

[Andyyyy64]

This is fixed in 0.5.18. I compared 0.5.17 and 0.5.18 using the same captured models and scores with simulated M2 Max 96 GB hardware. Qwen3.8-27B was absent from the old top ten and appeared second with the fix.

The grouping now preserves checkpoint versions and organization identity. Thanks for the detailed report.