Another small step for versioning, one giant leap for smooth inference 😎

🌍 1. New Model: Translategemma:4b

Tag: translategemma:4b

Break down language barriers effortlessly.
A powerful translation model designed to handle multilingual tasks with precision and speed — so you can translate text seamlessly and keep your workflows moving 🗣️

📖 Check out the model card for best practices and recommended use cases.

🩺 2. New Model: medgemma1.5:4b

Tag: medgemma1.5:4b

Your lightweight medical AI assistant.
A specialized model fine-tuned for the healthcare and medical domain. Perfect for quickly processing medical text, summarizing clinical notes, and extracting insights.

⚠️ Disclaimer

This tool (MedGemma + FastFlowLM) is not a diagnostic or clinical tool.
Always consult a licensed medical professional for healthcare decisions.

🖼️ 3. Auto Image Resize (VL Models)

Qwen family Vision models just got smarter about image size.

• Applies to: qwen2vl-it, qwen3vl-it, qwen2.5vl-it
• 📐 All input images are automatically resized to 1080p by default
• 🔧 Disable with setting arg -r 0:

flm run qwen2.5vl-it -r 0

🔒 Stability Improvements

• Increased stability for:

  • qwen3vl-it
  • qwen2.5vl-it

Fewer surprises. Smoother runs. More reliable vision inference 💪

As always, thank you for using FastFlowLM!
More speed. More models. More flow. 🚀

🖼️ New VLM Model Support

• Added: qwen2.5vl-it:3b – instruction-tuned VLM and ready to go!
• FYI: This model is based on Qwen2.5-VL-3B-Instruct, currently the most popular vision-text-to-text model measured by monthly downloads on Hugging Face! 🌍📥 (HF model leaderboard)

— The FastFlowLM Team

🧠 New Model Support

• Added: Qwen2.5-it:3b – instruction-tuned and ready to go!
• FYI: This model is based on Qwen2.5-3B-Instruct, currently the most popular text generation model measured by monthly downloads on Hugging Face! 🌍📥
  (Proof: HF model leaderboard)

⚡ Performance Improvements

• gpt-oss and gpt-oss-sg got a turbo boost! :race_car:
  • Up to 2.3× faster prefill speed for both short and long contexts.
  • Real numbers, real fast: Benchmark results here 📊

🐛 Bug Fixes

• Fixed an issue in medgemma
  • It's now fully squashed. You’re welcome 🪲🔨

That’s all for v0.9.32!
Stay fast, flow strong 💨
— The FastFlowLM Team

📦 What’s New

⚡ Prefill Speed Upgrade (Qwen & Gemma Families)

A new attention engine dramatically accelerates prefill, with larger gains at longer context lengths (especially 16K+).

• 🚀 Up to 3.8× faster prefill
  (qwen3:0.6b with a 32K-token prompt)

📈 Prefill Speed @ 32K Prompt (tokens/sec)

🖼️ Note: For qwen3vl-it:4b, image understanding is also faster in this release.

🔗 Benchmark Results

• 📊 Gemma3 performance: https://fastflowlm.com/docs/benchmarks/gemma3_results/
• 📈 Qwen3 performance: https://fastflowlm.com/docs/benchmarks/qwen3_results/

🛠️ Vision Tool Calling (New)

• ✅ Tool calling is now supported on qwen3vl-it
• 🔍 Enables vision tool calling workflows
• 🎥 Demo: https://youtu.be/Rf6r0Fm1UVs?si=u45hBgFXyDeEKXxh

🔧 Client Compatibility Improvements

• Non-stream mode logic adjusted
• Improves compatibility with client applications

🔧 What’s Fixed

🧠 GPT‑OSS Compatibility Bug

Patched a glitch that was causing hiccups with GPT‑OSS models.

Stay bug-free and model-happy 🐞🤖
— The FastFlowLM Team

⚠️ A bug affecting the gpt-oss:20b model was fixed in v0.9.30. Please use v0.9.30 or later.

📦 What’s New

⚡ Massive Prefill Speed Upgrade

We introduced a new attention engine that dramatically accelerates prefill, with larger gains at longer context lengths (especially 16K+).

• Up to 2.6× faster prefill
• More speedup at longer prompts
• No model re-download required (seamless upgrade)

📈 Prefill Speed with 32K prompt (tok/s)

🔜 Prefill speed upgrades for other models are on the way too — stay tuned!

📊 Detailed Benchmarks

• 🔗 LFM2 Results
• 🔗 LLaMA3 Results
• 🔗 Phi-4 Results

🩹 “No Pain” Update 😎

Just update and run — long-context prefill is now much faster. 🚀

🛠️ Tool Call Bug Fix (Non‑Stream Case)

We’ve squashed a pesky bug affecting tool calls when not using streaming.
Now it behaves exactly like you thought it should. 😌

Thanks for being awesome!
Happy modeling 🤖💙

— The FastFlowLM Team

Pain-free = no model redownload required to enjoy the speedup. Just update and go.

⚡️ 1. New Attention Engine — Gemma3 Vision Turbo

We’ve upgraded the Attention Engine to supercharge vision understanding for Gemma3-based models:

• Gemma3-4B
• MedGemma3-4B

📉 Latency Drop:
From 3.4s → 2.6s, giving you a ~25% speedup.
Your VLMs just got a whole lot snappier.

🖼️ 2. Qwen3-VL-Instruct-4B Gets a Vision Head Boost

Our shiny new Attention Engine also powers up the vision head for Qwen3-VL-Instruct-4B.
It really shines at high image resolutions.

🏎️ Example:
On 4K images, vision understanding is now ~45% faster compared to previous releases.
Less waiting, more interpreting.

🧘 More pain-free speedups coming soon — stay tuned!

✅ Key Benefits

• ✅ No redownloads — instant upgrade ⚡
• ✅ Lower vision latency 🕶️
• ✅ Better efficiency for VLM workloads 💡

🔗 Day-0 Support for LFM2.5-1.2B‑Thinking

First reasoning model from LiquidAI! More details: https://huggingface.co/LiquidAI/LFM2.5-1.2B-Thinking

LFM2.5-1.2B-Thinking delivers strong improvements in math reasoning, instruction following, and tool use, matching or exceeding Qwen3-1.7B on most reasoning benchmarks despite using ~40% fewer parameters

Model tag in FLM: lfm2.5-tk:1.2b

Run it in CLI Mode with:

flm run lfm2.5-tk:1.2b

Run it in Server Mode with:

flm serve lfm2.5-tk:1.2b

📊 Performance at a Glance

Kraken (Ryzen AI 340 / 350)

Decoding Speed vs. context length:

• 59 tok/s @ 4K context
• 52 tok/s @ 16K context

Strix/Strix Halo (Ryzen AI 360 and above)

Decoding Speed vs. context length:

• 54 tok/s @ 4K context
• 49 tok/s @ 16K context

Detailed benchmarks: https://fastflowlm.com/docs/benchmarks/lfm2_results/

Demo: 💧LFM2.5-1.2B-Thinking (LiquidAI) — 100% Powered by AMD Ryzen™ AI NPU

🆕 Model Tag Change

To avoid confusion between variants:

🧠 1. New Model: LiquidAI/LFM2‑2.6B‑Transcript

Tag: lfm2‑trans:2.6b

Summarize conference notes like a pro.
A single‑turn model designed to cleanly condense long transcripts into insights — so you can spend more time sipping ☕ and less time scrolling 📜.

🎬 See it in action: https://youtu.be/hpt0EhR1_vE?si=v9OCKa7VKAzuZ-02

🛠️ 2. Tool Calling — Preview Release

Tool calling is now officially out of preview!
Verified to work with:

• qwen3:4b
• qwen3:8b
• qwen3-it:4b
• qwen3-tk-4b

📹 Watch the demo: https://youtu.be/H-i4dztSdVk?si=5keyfkHt3ii8Wlu0

📘 Setup instructions (local):
👉 https://fastflowlm.com/docs/instructions/server/tool_calling/

💽 3. Installer Upgrade — xclbins Inside!

All xclbins are now bundled in the installer, which means:

• 🆙 Faster updates, No re‑downloading models unless weights change
• 🤯 Fewer user headaches
• 🚀 We are able to keep pushing the performance and efficiency.

More performance tuners coming soon… 🔧⚡

🔁 4. Runtime Restructure for Fine‑Tuned Models

We’ve overhauled the FastFlowLM runtime to let YOU plug in fine‑tuned models from supported families.

This is made possible by the upcoming gguf → q4nx conversion tool —
it’s almost ready and the docs are currently baking 🍳.

Stay tuned — this one will unlock a lot of flexibility.

🙌 Acknowledgements

• Huge thanks to @ItzCrazyKns from Perplexica for schooling us in the basics of tool calling and all the help along the way!
• Huge thanks to @jeremyfowers for highlighting and helping us resolving the ambiguity in the JSON-formatted reasoning output!

We're excited to introduce FastFlowLM v0.9.25, marking a key milestone with the integration of the new LFM2.5 model, freshly unveiled at CES 2026 (Jan 5th). This release also includes improvements to API compatibility and instruction-style models.

🚀 New Model Support

• LFM2.5-1.2B-Instruct
  🔸 Debuted at CES2026
  The newest addition to the LFM family, tuned for instruction-following. It features improved responsiveness and latency, ideal for interactive applications on AMD NPU.

• Phi4-mini-instruct
  A compact instruction model tailored for devices with limited memory — great for summarization and low-resource tasks.

🛠️ Fixes & Improvements

• ✅ Fixed bugs related to generation parameters (top_k, top_p, etc.) not being respected in OpenAI-compatible REST APIs.
• Ensures correct behavior when adjusting generation strategy through API calls.

🎉 With this update, FastFlowLM continues its mission to support the latest LLMs and provide an efficient, private, and developer-friendly experience on AMD Ryzen AI NPUs.