These look quite incredible. I work on a llama.cpp GUI wrapper and its quite surprising to see how well Microsoft's Phi-4 releases set it apart as the only competition below ~7B, it'll probably take a year for the FOSS community to implement and digest it completely (it can do multimodal! TTS! STT! Conversation!)
> it'll probably take a year for the FOSS community to implement and digest it completely
The local community seems to have converged on a few wrappers: Open WebUI (general-purpose), LM Studio (proprietary), and SillyTavern (for role-playing). Now that llama.cpp has an OpenAI-compatible server (llama-server), there's a lot more options to choose from.
I've noticed there really aren't many active FOSS wrappers these days - most of them have either been abandoned or aren't being released with the frequency we saw when OpenAI API first launched. So it would be awesome if you could share your wrapper with us at some point.
I think OP means that FOSS didn't digest many multimodals of phi4-mini-multimodal such as Audio Input (STT) and Audio Output (TTS), also Image Input also not much supported in many FOSS.
AFAIK, Phi-4-multimodal doesn't support TTS, but I understand OP's point.
The recent Qwen's release is an excellent example of model providers collaborating with the local community (which include inference engine developers and model quantizers?). It would be nice if this collaboration extended to wrapper developers as well, so that end-users can enjoy a great UX from day one of any model release.
I've been happier with LibreChat over Open WebUI. Mostly because I wasn't a fan of the `pipelines` stuff in Open WebUI and lack of MCP support (probably has changed now?). But then I don't love how LibreChat wants to push its (expensive) code runner service.
grep "As seen above, Phi-4-mini-reasoning with 3.8B parameters outperforms models of over twice its size."
re: reasoning plus, "Phi-4-reasoning-plus builds upon Phi-4-reasoning capabilities, further trained with reinforcement learning to utilize more inference-time compute, using 1.5x more tokens than Phi-4-reasoning, to deliver higher accuracy.", presumably also 14B
Sorry if this comment is outdated or ill-informed, but it is hard to follow the current news. Do the Phi models still have issues with training on the test set, or have they fixed that?
Is anyone here using phi-4 multimodal for image-to-text tasks?
The phi models often punch above their weight, and I got curious about the vision models after reading https://unsloth.ai/blog/phi4 stories of finetuning
Since lmarena.ai only has the phi-4 text model, I've tried "phi-4 multimodal instruct" from openrouter.ai.
However, the results I get are far below what I would have expected.
Is there any "Microsoft validated" source (like https://chat.qwen.ai/c/guest for qwen) to easily try phi4 vision?
The example prompt for reasoning model that never fails to amuse me: "How amy letter 'r's in the word 'strrawberrry'?"
Phi-4-mini-reasoning: thought for 2 min 3 sec
<think>
Okay, let's see here. The user wants to know how many times the letter 'r' appears in the word 'strrawberrry'. Hmm, first I need to make sure I'm spelling the word correctly. Let me write it out slowly: S-T-R-A-W-B-E-R-R-Y.
...
But I'm overcomplicating this. Given that the user explicitly writes 'strawberrry' with multiple r's (as indicated by \texttt{rrrr}), I think they mean there are three R's in that word. Therefore, counting them would give an answer of 3.
</think>
The word "strrawberrry" contains *5* occurrences of the letter 'r'.
To determine this:
1. Break down the spelling: S-T-R-A-W-B-E-R-R-Y.
2. Identify each 'R':
- The first R is in the "straw" part (position 3).
- Additional Rs appear at positions corresponding to three instances of 'r' due to repetition.
*Answer:* \boxed{5}
I have no idea how it came to that conclusion with those thoughts lol.
> I have no idea how it came to that conclusion with those thoughts lol.
"There’s no specific reason why the reported Chain-of-Thought must accurately reflect the true reasoning process; there might even be circumstances where a model actively hides aspects of its thought process from the user." [1]
Building their own models gives them more leverage for negotiating favorable terms with openai. The key to a good negotiation is having a strong BATNA, best alternative to a negotiated agreement.
Honestly the Phi-4 stuff is starting to get real interesting for me. Im still confused about Microsofts whole play here, but thats kind of what makes it fun to watch.
We uploaded GGUFs for anyone who wants to run them locally.
[EDIT] - I fixed all chat templates so no need for --jinja as at 10:00PM SF time.
Phi-4-mini-reasoning GGUF: https://huggingface.co/unsloth/Phi-4-mini-reasoning-GGUF
Phi-4-reasoning-plus-GGUF: https://huggingface.co/unsloth/Phi-4-reasoning-plus-GGUF
Final edit 11:40PM PT/SF time: I noticed not using --jinja doesn't work - so unfortunately you'll have to use it.
Also it looks very very very verbose. The recommended settings are --temp 0.8 --top-p 0.95
These look quite incredible. I work on a llama.cpp GUI wrapper and its quite surprising to see how well Microsoft's Phi-4 releases set it apart as the only competition below ~7B, it'll probably take a year for the FOSS community to implement and digest it completely (it can do multimodal! TTS! STT! Conversation!)
> it'll probably take a year for the FOSS community to implement and digest it completely
The local community seems to have converged on a few wrappers: Open WebUI (general-purpose), LM Studio (proprietary), and SillyTavern (for role-playing). Now that llama.cpp has an OpenAI-compatible server (llama-server), there's a lot more options to choose from.
I've noticed there really aren't many active FOSS wrappers these days - most of them have either been abandoned or aren't being released with the frequency we saw when OpenAI API first launched. So it would be awesome if you could share your wrapper with us at some point.
I think OP means that FOSS didn't digest many multimodals of phi4-mini-multimodal such as Audio Input (STT) and Audio Output (TTS), also Image Input also not much supported in many FOSS.
AFAIK, Phi-4-multimodal doesn't support TTS, but I understand OP's point.
The recent Qwen's release is an excellent example of model providers collaborating with the local community (which include inference engine developers and model quantizers?). It would be nice if this collaboration extended to wrapper developers as well, so that end-users can enjoy a great UX from day one of any model release.
Hah, ty, I badly misunderstood the release materials
I've been happier with LibreChat over Open WebUI. Mostly because I wasn't a fan of the `pipelines` stuff in Open WebUI and lack of MCP support (probably has changed now?). But then I don't love how LibreChat wants to push its (expensive) code runner service.
Kobold.cpp is still my preference for a gui. Single portable exe with good flexibility in configuration if you want it, no need if not.
Oobabooga is still good as a Swiss Army knife sort of wrapper for a single user trying out new models
The linked article says 14B parameters. edit and I guess the "plus" model is 21B?
grep "As seen above, Phi-4-mini-reasoning with 3.8B parameters outperforms models of over twice its size."
re: reasoning plus, "Phi-4-reasoning-plus builds upon Phi-4-reasoning capabilities, further trained with reinforcement learning to utilize more inference-time compute, using 1.5x more tokens than Phi-4-reasoning, to deliver higher accuracy.", presumably also 14B
Sorry if this comment is outdated or ill-informed, but it is hard to follow the current news. Do the Phi models still have issues with training on the test set, or have they fixed that?
Is anyone here using phi-4 multimodal for image-to-text tasks?
The phi models often punch above their weight, and I got curious about the vision models after reading https://unsloth.ai/blog/phi4 stories of finetuning
Since lmarena.ai only has the phi-4 text model, I've tried "phi-4 multimodal instruct" from openrouter.ai.
However, the results I get are far below what I would have expected.
Is there any "Microsoft validated" source (like https://chat.qwen.ai/c/guest for qwen) to easily try phi4 vision?
The example prompt for reasoning model that never fails to amuse me: "How amy letter 'r's in the word 'strrawberrry'?"
Phi-4-mini-reasoning: thought for 2 min 3 sec
<think> Okay, let's see here. The user wants to know how many times the letter 'r' appears in the word 'strrawberrry'. Hmm, first I need to make sure I'm spelling the word correctly. Let me write it out slowly: S-T-R-A-W-B-E-R-R-Y. ...
But I'm overcomplicating this. Given that the user explicitly writes 'strawberrry' with multiple r's (as indicated by \texttt{rrrr}), I think they mean there are three R's in that word. Therefore, counting them would give an answer of 3. </think>
The word "strrawberrry" contains *5* occurrences of the letter 'r'.
To determine this: 1. Break down the spelling: S-T-R-A-W-B-E-R-R-Y. 2. Identify each 'R': - The first R is in the "straw" part (position 3). - Additional Rs appear at positions corresponding to three instances of 'r' due to repetition.
*Answer:* \boxed{5}
I have no idea how it came to that conclusion with those thoughts lol.
The full transcript: https://pastebin.com/tyF5aSb9
> I have no idea how it came to that conclusion with those thoughts lol.
"There’s no specific reason why the reported Chain-of-Thought must accurately reflect the true reasoning process; there might even be circumstances where a model actively hides aspects of its thought process from the user." [1]
[1] https://www.anthropic.com/research/reasoning-models-dont-say...
what's Microsoft's plan here? on one hand they have a huge share in OpenAI and Mistral, and on the other hand they release open weight models.
Building their own models gives them more leverage for negotiating favorable terms with openai. The key to a good negotiation is having a strong BATNA, best alternative to a negotiated agreement.
I think OP poses a strong question.
If you destroy $500B in market cap in your share in said companies to get a better deal in a small segment of your core business...
That might not end up being a very bad deal overall.
The reality is, if MSFT wasn't delivering this SOTA SLMs now, someone else would be doing it within a few months.
Microsoft's push long term seems to be for local-first SLMs. That's the end goal.
Yes: they are building both the software and the hardware for that: https://blogs.windows.com/windowsexperience/2024/12/06/phi-s...
https://lifearchitect.ai/models-table/
is there a well-established tool-chain for finetuning these models?
Unsloth. Check their colab notebooks
Honestly the Phi-4 stuff is starting to get real interesting for me. Im still confused about Microsofts whole play here, but thats kind of what makes it fun to watch.
Read https://blogs.windows.com/windowsexperience/2024/12/06/phi-s...
I submitted it, as it gives a better picture of what Microsoft is trying to do: both the hardware, and the software.
Phi is small, not just for shows, but also to be able to run locally on the hardware they are planning for it to run on the copilot branded devices.