Menlo Research
Model fails to load/run in chat on Jan (v0.6.2 & v0.6.3) on Apple Silicon (M4) #5631
-
|
Hello everyone, I'm founding a persistent and annoying issue with the Jan app on my Apple Silicon Mac and was hoping to find out if this is a known bug or if there are any further troubleshooting steps. System specs: Hardware: MacBook Air (2024) with Apple M4 chip, 16GB RAM OS: macOS Sequoia 15.5 App versions: Jan.ai (tested with stable v0.6.3 from the website, v0.6.2 from GitHub Releases and 0.6.0 from website) Issue: Events: Initially, I tried importing a Gemma-3n-E4B-it-Q4_K_M.gguf model from a local folder (it runs smoothly on LM Studio). The Jan UI showed a successful import message, but the model didn't appear in the model list. I performed a complete clean install (deleting the app using AppCleaner and the ~/Library/Application Support/jan folder). After reinstalling (v0.6.2), I downloaded a different model (Sailor) directly through the Jan Hub. The model now appears correctly in the model list, and I can select it and "Start" it. However, when I go to a new chat and send my first message, the model fails to generate a response. The "thinking" indicator appears for a moment, then stops and shows an error message. When I check the model list again, the model is shown as "Stopped". This exact failure pattern occurs with EVERY SINGLE model I've tried, including Gemma-3n, Gemma3, DeepSeek R1, Sailor and models downloaded directly from the Jan Hub. This confirms the issue is not with the model files themselves. Troubleshooting steps already taken: Confirmed the model file is on the local SSD, not a cloud-synced folder (Dropbox). Performed a FULL "purge" of Jan, deleting all files from ~/Library/Application Support/jan/, ~/Library/Caches/jan/, etc., followed by a system restart and a clean install. Tested with multiple stable and older versions of Jan (v0.6.3, v0.6.2 and v0.6.0). Tested with multiple different GGUF models from different sources. This appears to be a fundamental bug in how the Jan application's inference engine interacts with the macOS/Apple Silicon (M4) environment. The app can list and prepare the model, but it fails on the first execution call. IT IS BROKEN. Is this a known issue? Are there any workarounds I've missed, or is this being addressed in a future update? I don't want to be rude but if it's not broken, don't fix it. This latest update removed a lot of features (model system) to make it more friendly but it was a DISASTER. Thanks for any help you can provide. |
Beta Was this translation helpful? Give feedback.
Replies: 11 comments 2 replies
-
|
Hi @rodrigobhz sorry for the frustration, just to check, have you tried this fix listed here yet? |
Beta Was this translation helpful? Give feedback.
-
|
Hi Bui,
Thanks for the fast reply.
I believe you may have misread my issue report. This isn't about models that stopped working after an update - it's about Jan not working at all on my system. I clearly stated that I tested 4 different releases (all 0.6+) and **none** of them can load or run models - whether downloaded from Jan's hub or loaded locally. This is a complete functionality failure, not a regression. Your response suggests troubleshooting steps for users whose Jan was previously working, which doesn't apply here. Could you provide guidance for cases where Jan has never successfully loaded models?
Best,
Rodrigo
… On 30 Jun 2025, at 22:14, Bui Quang Huy ***@***.***> wrote:
Hi @rodrigobhz <https://github.com/rodrigobhz> sorry for the frustration, just to check, have you tried this fix listed here yet?
#5564 <#5564>
—
Reply to this email directly, view it on GitHub <#5631 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BPVJ7H2MKMCP6HDKKPSKWN33GHOHLAVCNFSM6AAAAACAPNZZZKVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGNRSGMZTAOA>.
You are receiving this because you were mentioned.
|
Beta Was this translation helpful? Give feedback.
-
|
Hi @rodrigobhz, can you try to uninstall and do a fresh install of our latest v0.6.4 update? If that still doesn't work for you, please help provide the Thanks for your patience |
Beta Was this translation helpful? Give feedback.
-
|
[app.log](https://github.com/user-attachments/files/21028088/app.log)
[cortex.log](https://github.com/user-attachments/files/21028090/cortex.log)
Hi Bui,
Hope you're doing well.
It didn’t work at all. I did what you asked: uninstalled Jan, rebooted the machine, and installed the latest version with the hotfix, but the problems remain.
Here’s a detailed breakdown of what I've encountered:
I attempted to load gemma-3n-E4B-it-Q4_K_M.gguf and gemma-3n-E4B-it-Q3_K_M.gguf models locally. The llama.cpp modal indicates successful loading, but these models do not appear in the Jan.ai model list.
Similarly, I downloaded the Qwen3 and Gemma3 models directly from the Jan.ai hub, and they also fail to be listed.
The Phi-4-reasoning model did appear in the list, but initially, the chat showed a "model did not respond" error. I managed to resolve this by manually reducing the max_tokens in the model settings from 8192 to 4096. It worked when I started a new chat. Unfortunately, this workaround isn't possible for the unlisted models as I cannot access their parameters.
I've confirmed that all locally loaded/downloaded models are present in the Jan.ai models folder with their respective .yml files. I tested the integrity of the files and downloaded the local files from HugginFace. Both gemma-3n model files are working on LM Studio.
I've attached the Jan.ai logs (“app.log" and “cortex.log") for your review. Upon a quick inspection, I noticed recurring errors like "Failed to read config file," "Cortex server terminated unexpectedly," and "HTTP request failed with status code: 503" when the models attempt to load or start. There are also errors indicating "Value is not convertible to float" when updating the model.
I’m using a Macbook Air M4 16gb RAM running Sequoia 15.5.
Any insights or suggestions you might have would be greatly appreciated.
Thanks,
Rodrigo Ferreira

… On 2 Jul 2025, at 22:19, Bui Quang Huy ***@***.***> wrote:
Hi @rodrigobhz <https://github.com/rodrigobhz>, can you try to uninstall and do a fresh install of our latest v0.6.4 update? If that still doesn't work for you, please help provide the app log and cortex log from Settting -> General -> App log so we can inspect what might have been the problem.
Thanks for your patience
—
Reply to this email directly, view it on GitHub <#5631 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BPVJ7H3WVX6H4TDBK3ENNAD3GSAK5AVCNFSM6AAAAACAPNZZZKVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGNRUGY2TONQ>.
You are receiving this because you were mentioned.
|
Beta Was this translation helpful? Give feedback.
-
|
Hi @rodrigobhz, that’s such an annoying issue with Cortex where dangling models can’t be reimported. It might show as if you imported an existing model, but it’s not actually there (having been removed), though the symlink remains. Also, make sure to update the app to 0.6.4, as there’s a model load issue in 0.6.3 that blocks most model run requests. Try doing a factory reset and importing a smaller model (gemma3:1b, qwen3 0.6b) to see if they work. I believe gemma-3n isn’t supported yet. |
Beta Was this translation helpful? Give feedback.
-
|
Hi Louis/Bui,Thanks for the quick response. I appreciate your insights.However, I need to clarify that most of the suggested steps have already been thoroughly executed as detailed in my initial message and further observations. * App Version: I am already running Jan.ai version 0.6.4, which includes the latest hotfix. The issue persists on this version. Furthermore, I have tested all versions down to 0.6, including a nightly build, and the problem remained consistent across all of them. * Model Loading Issue (0.6.3 vs 0.6.4): My problems are occurring on 0.6.4, which suggests it's not solely the 0.6.3 specific bug. * "Dangling Models" and Reimporting: I performed a full reinstallation of Jan.ai and a reboot of my M4. This is more comprehensive than a factory reset within the app, as it ensures all application data and symlinks are cleared, and the problem still reoccurs immediately after a fresh install. * Testing Smaller Models: I have explicitly tested smaller models, including: * gemma3:1b (downloaded directly from Jan.ai hub) * qwen3 0.6b (downloaded directly from Jan.ai hub) * Sailor (approx. 750MB, also downloaded directly from Jan.ai hub) None of these models were listed in the Jan.ai interface after successful download/loading. This strongly indicates the issue is not related to model size or resource consumption for larger models, but rather a fundamental problem with how Jan.ai indexes and lists any model. * gemma-3n Compatibility: While gemma-3n was indeed one of the models I attempted to load (from a local file), the core issue extends to models downloaded directly from the Jan.ai hub (Qwen3, Gemma3, Sailor). If models from your own hub are not listing, then model-specific compatibility for gemma-3n is secondary to a broader indexing or UI display problem. It's important to note that upon my very first installation of Jan.ai, I successfully loaded and listed gemma3-n. However, when attempting to use it in the chat, I received a "model did not respond" error. This initial success, followed by the complete inability to list it again after file manipulation, suggests a regression or a state-dependent failure rather than a fundamental incompatibility.My logs consistently show [ERROR] Failed to run mcp commands: Failed to read config file and [WARN] Cortex server terminated unexpectedly, suggesting a deeper instability in the Jan.ai's backend or its interaction with the file system/model management layer, preventing models from being correctly listed or initialized even when their files are present and seemingly processed by llama.cpp.The only model that ever listed was Phi-4-reasoning, and only after I manually halved its max_tokens from 8192 to 4096, which implies a resource management sensitivity, but this workaround is impossible for models that never even appear in the list.The current behavior indicates a significant instability in Jan.ai's model management and loading pipeline on macOS M4.Let me know your thoughts based on this additional context.Best regards,Rodrigo FerreiraRodrigo FerreiraIllustrator + Graphic DesignerEm 3 de jul. de 2025, à(s) 00:44, Louis ***@***.***> escreveu:
Hi @rodrigobhz, that’s such an annoying issue with Cortex where dangling models can’t be reimported. It might show as if you imported an existing model, but it’s not actually there (having been removed), though the symlink remains. Also, make sure to update the app to 0.6.4, as there’s a model load issue in 0.6.3 that blocks most model run requests. Try doing a factory reset and importing a smaller model (gemma3:1b, qwen3 0.6b) to see if they work. I believe gemma-3n isn’t supported yet.
—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
Hi @rodrigobhz sorry for the frustration, I am going through the app log right now and here is some of my insight:
but it actually successfully restarted right after: We have noticed that there is this design problem in which the Hub page is cached, so if you click download models in the hub, sometimes it does so before the Cortex server spawned => it will hang the process afterward.
Here what I suggest you try. First do a full delete and uninstall of the app like you did in the first message:
Then, proceed to download Jan v0.6.4 Try to download a model from our hub first, before you first chat with it, set your context window size to a small number like 2048. If it is still failed, do let me know, and I will create a tickets to track this officially. |
Beta Was this translation helpful? Give feedback.
-
|
Hi Bui,Thanks for reviewing the logs. I appreciate your insights, but I need to clarify that the suggested steps have already been thoroughly and repeatedly executed. The issues persist, indicating a more complex problem.Here's a concise summary of my extensive testing: * Comprehensive Clean Installs: I have performed multiple full clean installations (including AppCleaner, manual ~/Library/Application Support/Jan folder deletion, and terminal commands to remove remnants) followed by system reboots. This was done for Jan v0.6.4, earlier versions down to 0.6.0, and a nightly build. * Hub Model Listing Failure: Critically, models downloaded directly from your official Jan.ai Hub (e.g., Qwen3 0.6b, Gemma3:1b, Sailor models) consistently fail to list in the UI on first download to a clean install. This rules out re-importation bugs for these specific cases. * Context Window / Parameters: I have not altered any general Jan.ai or llama.cpp parameters. The Phi-4-reasoning model was the only one that ever appeared in the list, requiring a manual max_tokens reduction (from 8192 to 4096) to function. For all other models, the problem is the inability to access these settings because they simply don't appear in the model list. * Initial Gemma3-n Behavior: On my very first Jan.ai installation, gemma3-n did load and list, but produced a "model did not respond" error in chat. After file manipulation, it ceased to list entirely, reinforcing a state-dependent issue. * Persistent Errors: My attached logs consistently show: * [WARN] Cortex server terminated unexpectedly and repeated restart attempts, with subsequent HTTP request failed with status code: 503 (communication failures with the local Cortex server). * Fail to import model, model_id 'model' already exists! or similar "already exists" errors for specific models, indicating indexing conflicts despite clean installs.The evidence points to a fundamental instability in Jan.ai's model management and loading pipeline on macOS M4, independent of user parameter settings.Please let me know if this clarifies the situation enough to open an official ticket.Best regards,Rodrigo FerreiraEm 3 de jul. de 2025, à(s) 13:24, Bui Quang Huy ***@***.***> escreveu:
Hi @rodrigobhz sorry for the frustration, I am going through the app log right now and here is some of my insight:
It seems that indeed your Cortex failed to start on the first attempt:
[2025-07-03][02:07:17][app_lib::core::setup][INFO] Cortex server spawned successfully. PID: 1580
[2025-07-03][02:07:17][app_lib::core::setup][INFO] [Cortex Terminated]: Signal Some(15), Code None
[2025-07-03][02:07:17][app_lib::core::setup][INFO] Cleared child process lock after termination.
[2025-07-03][02:07:17][app_lib::core::setup][WARN] Cortex server terminated unexpectedly.
[2025-07-03][02:07:17][app_lib::core::setup][INFO] Waiting 5000ms before attempting restart 1/5...
but it actually successfully restarted right after:
[2025-07-03][02:07:22][app_lib::core::setup][INFO] Spawning cortex-server (Attempt 2/5)
[2025-07-03][02:07:22][app_lib::core::setup][DEBUG] Sidecar command: Command { cmd: LD_LIBRARY_PATH=":/Applications/Jan.app/Contents/Resources/binaries" "/Applications/Jan.app/Contents/MacOS/cortex-server" "--start-server" "--port" "39291" "--config_file_path" "/Users/rodrigobhz/Library/Application Support/Jan/data/.janrc" "--data_folder_path" "/Users/rodrigobhz/Library/Application Support/Jan/data" "--cors" "ON" "--allowed_origins" "http://localhost:3000,http://localhost:1420,tauri://localhost,http://tauri.localhost" "config" "--api_keys" "U1HrjtLAaI6OrnoBxkXwA54e0E8mUcbu", raw_out: false }
[2025-07-03][02:07:22][app_lib::core::setup][INFO] Cortex server spawned successfully. PID: 1600
[2025-07-03][02:07:22][app_lib::core::setup][INFO] Cortex server started successfully, resetting restart count from 1 to 0.
[2025-07-03][02:09:09][reqwest::connect][DEBUG] starting new connection: http://127.0.0.1:39291/
[2025-07-03][02:10:03][reqwest::connect][DEBUG] starting new connection: http://127.0.0.1:39291/
We have noticed that there is this design problem in which the Hub page is cached, so if you click download models in the hub, sometimes it does so before the Cortex server spawned => it will hang the process afterward.
I think you might be having a combination of 2 different bug interfering with each other that cause the problem that you are facing:
one of them is this bug where if you imported a local model, delete them, and then re-import again. It will said that it is successful but it doesn't load. In this case to fix it, you will need to do a full factory reset.
the second problem is not a bug but rather your context window might have been set too high like what you experience with the Phi-4 model. The problem is that your model that you imported previously has ecountered the bug above, so changing the settings doesn't affect anything.
Here what I suggest you try. First do a full delete and uninstall of the app like you did in the first message:
I performed a complete clean install (deleting the app using AppCleaner and the ~/Library/Application Support/jan folder).
Then, proceed to download Jan v0.6.4
Try to download a model from our hub first, before you first chat with it, set your context window size to a small number like 2048.
If it is still failed, do let me know, and I will create a tickets to track this officially.
—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
Hi Bui/Louis,
Following our last exchange, I performed another full clean installation: AppCleaner, CleanMyMac, system reboot, then fresh Jan.ai v0.6.4 download and install.
The Phi-4-reasoning model, which I had previously used and manually adjusted its context window, appeared immediately active in the app's model list upon first launch of the freshly installed Jan.ai.
Despite being listed, attempting to interact with Phi-4-reasoning in the chat immediately resulted in the "model is not responding" modal. The model still listed on Jan interface.
This confirms a profound and concerning issue with Jan.ai's data persistence. Even after exhaustive cleaning methods, previous model data (including activation status) is inexplicably retained on the SSD, bypassing standard uninstallation procedures. This persistent "memory" of models, combined with their non-functional state upon re-detection, points to a severe underlying problem in Jan.ai's data management and Cortex initialization logic on macOS M4.
This behavior makes comprehensive debugging or achieving a true "clean state" for testing impossible at the user level. It suggests a deeply embedded cache or indexing system that survives aggressive purges.
Please advise on the next steps, as this issue seems to be beyond typical troubleshooting.
Best regards, Rodrigo Ferreira
… On 3 Jul 2025, at 13:24, Bui Quang Huy ***@***.***> wrote:
Hi @rodrigobhz <https://github.com/rodrigobhz> sorry for the frustration, I am going through the app log right now and here is some of my insight:
It seems that indeed your Cortex failed to start on the first attempt:
[2025-07-03][02:07:17][app_lib::core::setup][INFO] Cortex server spawned successfully. PID: 1580
[2025-07-03][02:07:17][app_lib::core::setup][INFO] [Cortex Terminated]: Signal Some(15), Code None
[2025-07-03][02:07:17][app_lib::core::setup][INFO] Cleared child process lock after termination.
[2025-07-03][02:07:17][app_lib::core::setup][WARN] Cortex server terminated unexpectedly.
[2025-07-03][02:07:17][app_lib::core::setup][INFO] Waiting 5000ms before attempting restart 1/5...
but it actually successfully restarted right after:
[2025-07-03][02:07:22][app_lib::core::setup][INFO] Spawning cortex-server (Attempt 2/5)
[2025-07-03][02:07:22][app_lib::core::setup][DEBUG] Sidecar command: Command { cmd: LD_LIBRARY_PATH=":/Applications/Jan.app/Contents/Resources/binaries" "/Applications/Jan.app/Contents/MacOS/cortex-server" "--start-server" "--port" "39291" "--config_file_path" "/Users/rodrigobhz/Library/Application Support/Jan/data/.janrc" "--data_folder_path" "/Users/rodrigobhz/Library/Application Support/Jan/data" "--cors" "ON" "--allowed_origins" "http://localhost:3000,http://localhost:1420,tauri://localhost,http://tauri.localhost" "config" "--api_keys" "U1HrjtLAaI6OrnoBxkXwA54e0E8mUcbu", raw_out: false }
[2025-07-03][02:07:22][app_lib::core::setup][INFO] Cortex server spawned successfully. PID: 1600
[2025-07-03][02:07:22][app_lib::core::setup][INFO] Cortex server started successfully, resetting restart count from 1 to 0.
[2025-07-03][02:09:09][reqwest::connect][DEBUG] starting new connection: http://127.0.0.1:39291/
[2025-07-03][02:10:03][reqwest::connect][DEBUG] starting new connection: http://127.0.0.1:39291/
We have noticed that there is this design problem in which the Hub page is cached, so if you click download models in the hub, sometimes it does so before the Cortex server spawned => it will hang the process afterward.
I think you might be having a combination of 2 different bug interfering with each other that cause the problem that you are facing:
one of them is this bug where if you imported a local model, delete them, and then re-import again. It will said that it is successful but it doesn't load. In this case to fix it, you will need to do a full factory reset.
the second problem is not a bug but rather your context window might have been set too high like what you experience with the Phi-4 model. The problem is that your model that you imported previously has ecountered the bug above, so changing the settings doesn't affect anything.
Here what I suggest you try. First do a full delete and uninstall of the app like you did in the first message:
I performed a complete clean install (deleting the app using AppCleaner and the ~/Library/Application Support/jan folder).
Then, proceed to download Jan v0.6.4
Try to download a model from our hub first, before you first chat with it, set your context window size to a small number like 2048.
If it is still failed, do let me know, and I will create a tickets to track this officially.
—
Reply to this email directly, view it on GitHub <#5631 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BPVJ7H4FOJ5RMEMBXUIEACD3GVKNHAVCNFSM6AAAAACAPNZZZKVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGNRVGQ3DIMA>.
You are receiving this because you were mentioned.
|
Beta Was this translation helpful? Give feedback.
-
|
Hi @rodrigobhz can you help us providing the latest |
Beta Was this translation helpful? Give feedback.
-
|
Hi Bui,
Yes, of course. Here are the latest app.log and cortex.log from my most recent testing session, following our last exchange and the full factory reset procedure.
To clarify the latest status after my previous emails:
I performed another full factory reset (using AppCleaner, CleanMyMac, and a system reboot), then reinstalled Jan.ai v0.6.4. This time, I observed improved model listing:
I successfully downloaded and listed Qwen 2.5 from the Jan.ai Hub.
I successfully loaded and listed Gemma-3n Q4 and Gemma-3n Q3 (local files).
However, the core issue of models not responding in chat persists for Gemma-3n, even with manual parameter adjustments:
Qwen 2.5 responded quickly after I reduced its max_tokens.
For both Gemma-3n Q4 and Q3, even after reducing max_tokens, they continued to produce the "model is not responding" error in chat.
I think maybe this persistent "model not responding" issue for Gemma-3n models stems from Jan.ai utilizing an outdated llama.cpp version. I've observed that LM Studio, which uses a more recent llama.cpp build, is capable of running these newer models without such issues on my M4. This suggests a compatibility gap between the bundled llama.cpp in Jan.ai and newer GGUF model architectures, leading to inference failures.
The logs attached correspond to these latest tests and observations.
Best regards, Rodrigo Ferreira
[app.log](https://github.com/user-attachments/files/21065032/app.log)
[cortex.log](https://github.com/user-attachments/files/21065033/cortex.log)

… On 3 Jul 2025, at 22:41, Bui Quang Huy ***@***.***> wrote:
Hi @rodrigobhz <https://github.com/rodrigobhz> can you help us providing the latest app log and cortex log again?
—
Reply to this email directly, view it on GitHub <#5631 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BPVJ7H6FQK4M57YAJCS4ALT3GXLS7AVCNFSM6AAAAACAPNZZZKVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGNRVG44TOMA>.
You are receiving this because you were mentioned.
|
Beta Was this translation helpful? Give feedback.
-
|
Hi @rodrigobhz I will mark this as answered now, the problem with Gemma-3n is known to us because right now we are transitioning from our old engine to fully llama.cpp. This problem will be corrected in v0.7.0, so do please be a bit patient with us. Do let us know if you still have problem with loading any other model. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks!Rodrigo FerreiraIllustrator + Graphic DesignerEm 4 de jul. de 2025, à(s) 23:37, Bui Quang Huy ***@***.***> escreveu:
Hi @rodrigobhz I will mark this as answered now, the problem with Gemma-3n is known to us because right now we are transitioning from our old engine to fully llama.cpp. This problem will be corrected in v0.7.0, so do please be a bit patient with us. Do let us know if you still have problem with loading any other model.
—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
Hi @rodrigobhz I will mark this as answered now, the problem with Gemma-3n is known to us because right now we are transitioning from our old engine to fully llama.cpp. This problem will be corrected in v0.7.0, so do please be a bit patient with us. Do let us know if you still have problem with loading any other model.