🧠 FastFlowLM v0.9.6 — More Qwen3 Variants & Penalty Logic Fix

FastFlowLM v0.9.6 introduces powerful new models and fine-tunes sampling behavior to match expected transformer behavior more accurately.

---

✨ What's New

🆕 1. New Model Support

• ✅ qwen3-tk:4b — Qwen3-4B Thinking model, supporting up to 256k context length (default 32768)
• ✅ qwen3-it:4b — Qwen3-4B Instruction-tuned model (no think mode), supporting up to 256k context length (default 32768)
• ✅ deepseek-r1-0528:8b — DeepSeek-R1 0528 distilled variant of Qwen3:8b, supporting up to 64k context length (default 16384)

These additions expand FastFlowLM's coverage of task-optimized and instruction-tuned models.

---

🛠️ 2. Penalty Handling Fixes

Behavior Update

• Now, both repetition_penalty and frequency_penalty:
  • Do nothing when set to 1.0 (consistent with HuggingFace Transformers).
  • Apply penalty when > 1.0
  • Apply reward (encouragement) when < 1.0, not recommended.

Defaults Updated

• The default values for both penalties are now set to:

  repetition_penalty = 1.1
  frequency_penalty = 1.1
  

This makes the decoding behavior clearer and more intuitive for advanced users.

---

✅ Usage Reminder

Instruction-tuned Qwen3:

flm run qwen3-it:4b

Thinking-only Qwen3:

flm run qwen3-tk:4b

DeepSeek R1 0528:

flm run deepseek-r1-0528:8b

---

This release improves flexibility in model selection and ensures better alignment with standard decoding semantics.