Having major issues loading in any models. even a small one

[saintrac]

Category

Model inference issue

What happened

I tried to load a small model after the update to see if everything got fixed for me and it seems that no model is loading in the Nativ app for me what so ever. It get's stuck at Loading 100% and doesn't do anything else at all. Basically not usable.

Diagnostics

Environment

• Nativ: 0.3.4 (202608241945)
• macOS: macOS 26.5.2 (Build 25F84)
• Chip: Apple M4 Max
• Memory: 64 GB total, 55.89 GB in use
• mlx-vlm: 0.6.16

Model & server

• Server: running
• Selected model: LiquidAI/LFM2.5-2.6B
• Loaded model: None
• Max output tokens: 2048
• Context window: model default
• KV quantization: off
• Speculative decoding: off
• Prefix caching: off
• Thinking: on
• Launch arguments: --host 0.0.0.0 --port 8080 --max-tokens 2048 --model LiquidAI/LFM2.5-2.6B

Inference

• Metrics unavailable: The resource could not be loaded because the App Transport Security policy requires the use of a secure connection.
• — All-time —
• Requests completed: 11
• Requests failed: 0
• Prompt tokens: 226K
• Generated tokens: 4.7K
• Total processed tokens: 230K
• Avg decode speed: 16.8 tok/s
• Avg request speed: 4.7 tok/s

Server output (last 33 lines)

Started mlx-vlm-server.
INFO:     Started server process [65590]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8080 (Press CTRL C to quit)
Stopping mlx-vlm-server...
INFO:     Shutting down
INFO:     Waiting for application shutdown.
INFO:     Application shutdown complete.
INFO:     Finished server process [65590]
mlx-vlm-server stopped after Nativ requested a model switch (status 143)
Started mlx-vlm-server.
INFO:     Started server process [66841]
INFO:     Waiting for application startup.
2026-08-25 06:42:12,223 - INFO - Pre-loading language model: LiquidAI/LFM2.5-2.6B
2026-08-25 06:42:12,227 - INFO - Loading model: LiquidAI/LFM2.5-2.6B
2026-08-25 06:42:12,532 - INFO - HTTP Request: GET https://huggingface.co/api/models/LiquidAI/LFM2.5-2.6B/revision/main "HTTP/1.1 200 OK"
2026-08-25 06:42:12,594 - INFO - HTTP Request: GET https://huggingface.co/api/models/LiquidAI/LFM2.5-2.6B/tree/654f9463ce32b05d0429d76fe1f580b27d4c1ac0?recursive=true&expand=false "HTTP/1.1 200 OK"
Fetching 8 files:   0%|          | 0/8 [00:00<?, ?it/s]2026-08-25 06:42:12,938 - INFO - HTTP Request: HEAD https://huggingface.co/LiquidAI/LFM2.5-2.6B/resolve/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/generation_config.json "HTTP/1.1 307 Temporary Redirect"
2026-08-25 06:42:12,996 - INFO - HTTP Request: HEAD https://huggingface.co/api/resolve-cache/models/LiquidAI/LFM2.5-2.6B/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/generation_config.json "HTTP/1.1 200 OK"
Fetching 8 files:  50%|█████     | 4/8 [00:00<00:00,  9.97it/s]2026-08-25 06:42:13,019 - INFO - HTTP Request: HEAD https://huggingface.co/LiquidAI/LFM2.5-2.6B/resolve/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/model.safetensors.index.json "HTTP/1.1 307 Temporary Redirect"
2026-08-25 06:42:13,026 - INFO - HTTP Request: HEAD https://huggingface.co/LiquidAI/LFM2.5-2.6B/resolve/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/tokenizer_config.json "HTTP/1.1 307 Temporary Redirect"
2026-08-25 06:42:13,027 - INFO - HTTP Request: HEAD https://huggingface.co/LiquidAI/LFM2.5-2.6B/resolve/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/chat_template.jinja "HTTP/1.1 307 Temporary Redirect"
2026-08-25 06:42:13,029 - INFO - HTTP Request: HEAD https://huggingface.co/LiquidAI/LFM2.5-2.6B/resolve/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/config.json "HTTP/1.1 307 Temporary Redirect"
2026-08-25 06:42:13,070 - INFO - HTTP Request: HEAD https://huggingface.co/api/resolve-cache/models/LiquidAI/LFM2.5-2.6B/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/model.safetensors.index.json "HTTP/1.1 200 OK"
2026-08-25 06:42:13,080 - INFO - HTTP Request: HEAD https://huggingface.co/api/resolve-cache/models/LiquidAI/LFM2.5-2.6B/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/chat_template.jinja "HTTP/1.1 200 OK"
2026-08-25 06:42:13,082 - INFO - HTTP Request: HEAD https://huggingface.co/api/resolve-cache/models/LiquidAI/LFM2.5-2.6B/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/tokenizer_config.json "HTTP/1.1 200 OK"
2026-08-25 06:42:13,085 - INFO - HTTP Request: HEAD https://huggingface.co/api/resolve-cache/models/LiquidAI/LFM2.5-2.6B/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/config.json "HTTP/1.1 200 OK"
Fetching 8 files: 100%|██████████| 8/8 [00:00<00:00, 16.32it/s]
2026-08-25 06:42:15,619 - INFO - Model and processor loaded successfully.
2026-08-25 06:42:15,620 - INFO - Language model ready, continuous batching enabled.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8080 (Press CTRL C to quit)

[linear-code]

NEY-74

[Lazarus-931]

hi @saintrac, thx for reporting issues.

I see you're your memory is 64 GB total, with 55.89 GB in use.

Do you get the same experience when you download with less memory being used?

[saintrac]

Yes I get the same experience. I even removed all local models that I had and even deleted and reinstalled Nativ but still it doesn't work. @Lazarus-931 . I also saw that on my setting "See attached screenshot". I tried pressing try again but it doesn't do anything. I also still have issues downloading the Minimax-Music3 model. I don't know if it's the same reason why.

[Yavaren]

@saintrac Has this issue always been present, or has it appeared after a certain release?

[ivanfioravanti]

I have same issue on a machine and I have been able to create repro steps. Problem is when 8080 ports is used by another service. In my case nginx. /cc @Lazarus-931 @Yavaren it's present since initial release.

This is

Reproduction

1. Quit Nativ.
2. Start Homebrew nginx, whose default configuration listens on port 8080:

brew install nginx
brew services start nginx
curl -i http://127.0.0.1:8080/metrics

Confirm the response is 404 Not Found from nginx.

3. Open Nativ v0.3.5.
4. Keep the server configured as:

Host: 127.0.0.1
Port: 8080

5. Select a downloaded language model and start the server.
6. Wait until Nativ’s server logs report that the model loaded.
7. Observe that the UI remains at “Loading model”.

Confirm the collision:

lsof -nP -iTCP:8080 -sTCP:LISTEN

Both nginx and Nativ’s Python process may appear.

Confirm where UI polling goes:

tail -f /opt/homebrew/var/log/nginx/access.log

You should see repeated requests resembling:

GET /metrics HTTP/1.1 404
GET /v1/metrics HTTP/1.1 404

Ready-to-paste issue

Title: Metrics polling failures leave the UI stuck in “Loading model”

Environment:

macOS 27.0
Nativ 0.3.5
Server host: 127.0.0.1
Server port: 8080
Competing service: nginx listening on *:8080

Expected behavior:

Nativ should refuse to start when the configured port is occupied. If metrics polling subsequently fails with a non-transient error, the UI should leave the loading state and display an endpoint or port-conflict error.

Actual behavior:

The model server reports that the model loaded, but /metrics and /v1/metrics requests reach nginx and return HTTP 404. Nativ retries indefinitely and the UI remains in “Loading model”.

Probable cause:

handleMetricsFetchFailure records lastMetricsError, but does not clear metricsLoading, modelLoadingProgress, or the model-switch state. Only a successful metrics fetch clears them.

Relevant code:

• Sources/Nativ/NativModel.swift:260
• Sources/Nativ/NativModel.swift:928
• Sources/Nativ/NativModel.swift:960

Suggested fix:

After the startup grace period, a definitive error—especially when both metrics endpoints return 404—should:

• Exit the loading state.
• Unlock model controls.
• Surface an endpoint/port-conflict error.
• Continue lower-frequency health checks if automatic recovery is desired.

Mention that commit c1d5a3b already adds startup port detection, but the metrics-failure fallback should still be hardened.

If nginx was installed only for this test:

brew services stop nginx

[Lazarus-931]

hey @ivanfioravanti, yep that issue was fixed

I have same issue on a machine and I have been able to create repro steps. Problem is when 8080 ports is used by another service. In my case nginx. /cc @Lazarus-931 @Yavaren it's present since initial release.

This is

Reproduction

1. Quit Nativ.
2. Start Homebrew nginx, whose default configuration listens on port 8080:

brew install nginx brew services start nginx curl -i http://127.0.0.1:8080/metrics

Confirm the response is 404 Not Found from nginx.

3. Open Nativ v0.3.5.
4. Keep the server configured as:

Host: 127.0.0.1 Port: 8080

5. Select a downloaded language model and start the server.
6. Wait until Nativ’s server logs report that the model loaded.
7. Observe that the UI remains at “Loading model”.

Confirm the collision:

lsof -nP -iTCP:8080 -sTCP:LISTEN

Both nginx and Nativ’s Python process may appear.

Confirm where UI polling goes:

tail -f /opt/homebrew/var/log/nginx/access.log

You should see repeated requests resembling:

GET /metrics HTTP/1.1 404 GET /v1/metrics HTTP/1.1 404

Ready-to-paste issue

Title: Metrics polling failures leave the UI stuck in “Loading model”

Environment:

macOS 27.0 Nativ 0.3.5 Server host: 127.0.0.1 Server port: 8080 Competing service: nginx listening on *:8080

Expected behavior:

Nativ should refuse to start when the configured port is occupied. If metrics polling subsequently fails with a non-transient error, the UI should leave the loading state and display an endpoint or port-conflict error.

Actual behavior:

The model server reports that the model loaded, but /metrics and /v1/metrics requests reach nginx and return HTTP 404. Nativ retries indefinitely and the UI remains in “Loading model”.

Probable cause:

handleMetricsFetchFailure records lastMetricsError, but does not clear metricsLoading, modelLoadingProgress, or the model-switch state. Only a successful metrics fetch clears them.

Relevant code:

• Sources/Nativ/NativModel.swift:260
• Sources/Nativ/NativModel.swift:928
• Sources/Nativ/NativModel.swift:960

Suggested fix:

After the startup grace period, a definitive error—especially when both metrics endpoints return 404—should:

• Exit the loading state.
• Unlock model controls.
• Surface an endpoint/port-conflict error.
• Continue lower-frequency health checks if automatic recovery is desired.

Mention that commit c1d5a3b already adds startup port detection, but the metrics-failure fallback should still be hardened.

If nginx was installed only for this test:

brew services stop nginx

hi @ivanfioravanti, thx for the detailed report :)

i believe his should be fixed with #406, can you try rebuilding from main?

Nativ should properly catch the error reported by mlx-vlm server.

Also, are you as well experiencing model loading outside of this port issue?

[ivanfioravanti]

Yes @Lazarus-931 🚀

[Lazarus-931]

hey @ivanfioravanti, do you still experience this on 0.3.7?