Flint / FreeToken Runpod result — 2026-10-23
Flint runs through FreeToken's open PR #797, pinned to 711a86ef11c67bbe6c69ff59370594f64c6ef1b8. This validates an isolated Linux runtime, not the installed Windows Desktop engine.
| Check | Official NVFP4 GGUF | Unsloth UD-Q4_K_XL |
|---|---|---|
| Reference/config/tokenizer/parser tests | 217 passed | 217 passed |
| Visible exact reply | Passed | Passed |
| Generated interval-merging code, five functional cases | Passed | Returned text failed Python parsing; unresolved |
| Tool-call/result round trip | Passed | Passed |
| Streaming | Passed | Passed |
| 32K profile retrieval | All three keys recovered from 29,528 input tokens | Not run after coding-case failure |
The NVFP4 retrieval script initially reported failure because it counted a tokenizer return object instead of token IDs. The saved server usage and visible answer independently establish the prompt length and successful recovery. The original failure flag remains in the raw evidence; this is a reviewed result, not a claimed clean test exit.
Environment: Runpod Secure RTX 5090 32 GB, CUDA host 13.2; PyTorch 2.14.1, Triton 3.8.0, Transformers 5.18.0. Base wheels match commit 507a8bfa8, with the PR's Python source and installed native modules. Both models used hybrid experts, 2 GiB expert cache, eight CPU workers and one concurrent request. NVFP4 used about 19.8 GiB host expert banks and had 26.67 GiB GPU memory free after 8K graph capture. Pod performance cannot be projected directly onto a 5070 Ti or Windows/WSL.
Initial attempts exposed probe readiness, orphaned-process and executable-PATH problems. Corrected runs produced the results above. No FreeToken implementation code or local default installation was changed. Q4's failed coding response was not retained by the first acceptance helper, so its cause cannot be diagnosed from the saved error alone.
Recommendation: use official NVFP4 as the first local integration candidate. Q4 is viable for loading, chat and tools, but needs a retained-response coding rerun before choosing it over NVFP4. The current installed engine lacks Flint registration; a separate candidate environment or a future release incorporating PR #797 is required.
Cleanup: pod c37e106tar2zv0 in EUR-NO-1 was terminated; the following list-pods returned an empty inventory. No separate network volume was created. Runtime approximately 34 minutes at $1.19/hour GPU plus estimated disk cost: approximately $0.70 total, not verified invoice billing. Image: runpod/pytorch:1.0.7-cu1300-torch291-ubuntu2404-cluster. No live endpoint remains.
Sources: https://github.com/FlashML-org/FreeToken/pull/797 and https://huggingface.co/RecursiveSI/flint-nvfp4-gguf. Raw evidence is flint-runpod-evidence.tar.gz alongside this report; it contains logs/results and no weights or credentials.