Description

During onboarding, the inference provider model selection screen shows model names
only — no file size, parameter count, quantization, or estimated VRAM requirement
is displayed before the user selects a model for download. Users on machines with
limited VRAM can unknowingly select a model that will OOM or fail to load, with
no in-UI signal that the model exceeds their GPU capacity.

Platform scope: Reproduced on WSL2 (Ubuntu 24.04, x86_64) only; other platforms
not tested.
Regression: Unknown — earlier versions not tested for this specific display.

Environment

Device:        Windows PC with NVIDIA RTX PRO 6000 Blackwell Server (97 GB)
OS:            Ubuntu 24.04.4 LTS (WSL2 under Windows)
Architecture:  x86_64
Node.js:       v22.23.1
npm:           10.9.8
Docker:        29.5.3
OpenShell CLI: 0.0.85
NemoClaw:      v0.0.93
OpenClaw:      2026.7.1 (2d2ddc4)

Steps to Reproduce

1. Run nemoclaw onboard (or nemoclaw onboard --resume after a failed express install).

2. At [3/8] Configuring inference provider, select a local Ollama option
   (e.g. "Local Ollama (Windows host:11434) — running (suggested)").

3. Observe the model selection screen:

   Ollama models:
     1) qwen3.6:35b
     2) Other...
   
   Choose model [1]:
   

4. Note: no file size, VRAM requirement, or headroom is shown alongside the model name.

5. Select "2) Other..." and type any model name — again, no size guidance is shown.

Expected Result

Model selection shows size metadata alongside each model name so users can make
an informed choice before committing to a potentially multi-GB download. For example:

Ollama models:
  1) qwen3.6:35b  (23 GB · ~20 GB VRAM · fits: 73 GB headroom on this GPU)
  2) Other...

Or at minimum, a note such as:

Available VRAM: 92.8 GB — select a model that fits your GPU

This is especially important for express install, where the model is chosen
non-interactively: the selected default should be validated against detected VRAM
and the user should be informed of the choice and its size.

Actual Result

Model selection shows names only:

Ollama models:
  1) qwen3.6:35b
  2) Other...

Choose model [1]:

No file size, parameter count, quantization, or VRAM requirement is shown.
Users on small-GPU machines (e.g. 8 GB VRAM) have no UI signal that a model
like qwen3.6:35b (requires ~20 GB VRAM) will fail to load.

The WSL Ollama startup log shows VRAM info internally
(available="92.8 GiB", vram-based default_num_ctx=262144) but this information
is not surfaced in the NemoClaw wizard.

Logs

WSL Ollama startup (internal VRAM detection — not shown to user):
  inference compute: CUDA0 NVIDIA RTX PRO 6000 Blackwell Server Edition
    total="95.3 GiB" available="92.8 GiB"
  vram-based default context: total_vram="95.3 GiB" default_num_ctx=262144

ollama show qwen3.5:9b (metadata available but not surfaced in wizard):
  parameters     9.7B
  context length 262144
  quantization   Q4_K_M

ollama show qwen3.6:35b:
  parameters     36.0B
  context length 262144
  quantization   Q4_K_M

Related

• [All Platforms] Show model size and prompt for confirmation before Ollama model download #2639 (closed) — "Show model size and prompt for confirmation before Ollama model download": earlier enhancement covering the download-confirmation step; this bug extends the scope to the model selection screen and VRAM headroom display.
• [WSL2][Inference] qwen3.6:35b yields no token within 300s under OpenClaw agent loop on ARM64/aarch64 (64 GB iGPU) #3707 (open) — "[WSL2][Inference] qwen3.6:35b yields no token within 300s": a downstream consequence of selecting a model without VRAM guidance.