Qwen 3.5

5 atoms · 跨 4 天 · 首见 2026-04-26 · 最近 2026-06-26

三色: 🟦 fact 3 · 🟥 take 2 · stance ▲1/▼0/◆3

来源: X 5

时态: fresh:5

标签: 好思考:2 · 好数字:1 · 好观点:1 · 好信源:1

叙事 (narrative) (1)

• 2026-06-26 · models always get stupider when i abliterate them, gotta do a repair training afterwards
  they do, gotta do a repair training afterwards
  tags: 好思考 · → daily

预测 (forecast) (1)

• 2026-05-08 · can be used as base model for fine-tuned SLM
  I'd charge them a $10k to $20k one-time fee. Use Qwen 3.5 or Gemma 4 as base models, use Codex as the brain and DeepSeek v4 + Kimi as the muscle, and post-train a strong SLM under $1000.
  tags: 好观点 · → daily

事实 (fact) (3)

• 2026-06-26 · whitebox repair tune using samples from big boi Qwen 3.5
  after ablating I did a whitebox repair tune using samples from big boi Qwen 3.5, on tasks like code and agent things
  tags: 好思考 · → daily
• 2026-05-29 · 使用了专家蒸馏技术
  All the recent LLM tech reports use some kind of expert distillation: Qwen 3.5
  tags: 好信源 · → daily
• 2026-04-26 · 27B模型中DeltaNet层占比
  75% of Qwen 3.5 27B layers are DeltaNet (linear attention) and not softmax / full attention
  tags: 好数字 · → daily
  value: qty=75% · date=2026-04-26

时间轴 (近 20)

• 2026-06-26 · narrative · models always get stupider when i abliterate them, gotta do · → daily
• 2026-06-26 · fact · whitebox repair tune using samples from big boi Qwen 3.5 · → daily
• 2026-05-29 · fact · 使用了专家蒸馏技术 · → daily
• 2026-05-08 · forecast · can be used as base model for fine-tuned SLM · → daily
• 2026-04-26 · fact · 27B模型中DeltaNet层占比 · → daily

← 实体目录 · 系统日志