Flint Fixed and Sharp Chat Templates
Template-only release for Recursive Flint. No model weights or quantizations are included.
This repository contains two standalone Jinja chat templates: a schema- and compatibility-focused Fixed variant and a more concise Fixed + Sharp variant. The files were evaluated in three single NanoBench v0.7.1 runs using the operator-attributed flint-ud-iq3_xxs configuration in OpenCode with a 65,536-token (64 Ki-token) context window. The context setting is operator-reported; NanoBench run metadata leaves the context field null. The exact quantization is not recorded.
Templates
| Variant | File | Behavior | Observed score |
|---|---|---|---|
| Flint Fixed | templates/flint_fixed.jinja |
Compatibility fixes without Sharp instructions. | 100.0 / 100 |
| Flint Fixed + Sharp | templates/flint_fixed_sharp.jinja |
Fixed behavior plus concise-answer guidance; terse is enabled by default. |
100.0 / 100 |
Both scores are single-run observations, not a guarantee of future performance. The original template scored 98.8 / 100 in one run; its only lost points were on coding-05.
Benchmark snapshot
NanoBench v0.7.1 evaluated 13 tasks under the same task-set SHA-256 7e5baefb6485699c93f1aaa67de0b642fa3b5ffb312086340e48347d7aef115f.
| Configuration | Score | Tasks solved | Official run time | Tool calls |
|---|---|---|---|---|
| Original template | 98.8 / 100 | 12 / 13 | 6,476.142 s | 378 |
| Fixed | 100.0 / 100 | 13 / 13 | 6,181.937 s | 368 |
| Fixed + Sharp | 100.0 / 100 | 13 / 13 | 5,153.719 s | 295 |
Fixed and Fixed + Sharp both received full points in their single runs. Sharp finished 16.6% faster than Fixed and made 19.8% fewer tool calls, with a similar output-token total. Its session also had more context compactions: nine versus six for Fixed. Compaction input raised its recorded input total to 670,349 tokens, so raw input totals should not be read as a direct measure of task work.
See the full benchmark report, session-level data, and per-task data.
Attribution and identity notes
I asked GPT Astra to audit and fix the Flint chat template, then adapt the peculiar-ragdoll Sharp configuration. The Fixed template is adapted from RecursiveSI's official Flint chat template. RecursiveSI's model card lists Apache-2.0 licensing. These Jinja files are modified community templates, not official RecursiveSI or peculiar-ragdoll releases.
The benchmark report labels the runs as flint-ud-iq3_xxs in OpenCode and reports a 65,536-token context window. NanoBench results are CLIENT_REPORTED; the quantization and model file are not independently verified by the runner, and its own context field is null. Public attribution has caveats: the original and Fixed records were corrected after their scores froze, while the Sharp public entry remains labelled current. These metadata issues do not change the frozen scores. Links and details are in the report.
Usage
For a compatible llama.cpp build with Flint support, pass the selected template file:
llama-server \
-m /path/to/flint.gguf \
--jinja \
--chat-template-file templates/flint_fixed.jinja
For Fixed + Sharp, change the final path to templates/flint_fixed_sharp.jinja. When the serving layer forwards chat_template_kwargs, set these values to turn off the added Sharp guidance:
{"terse": false, "tool_fidelity": false}
Tool-call arguments should be passed as JSON objects, not unparsed JSON strings. Compatibility was evaluated in the reported OpenCode setup only; this package does not claim validation across all parsers, runtimes, clients, or model files.
Limitations
- There is one run per variant. A 100 / 100 score in one run is not a repeatability estimate.
- NanoBench run artifacts do not save the exact chat-template SHA-256. The current source files were modified before the corresponding Fixed and Sharp runs, but the run records do not cryptographically identify those file bytes. The current hashes and stale-manifest discrepancy are documented in
release_manifest.jsonand the local pre-upload notes. - OpenCode database metrics are reported through the operator-supplied summary; the raw OpenCode database and private execution logs are not included.
- Raw input-token totals include context-compaction rereads and are not directly comparable as task-work counts. The Sharp session compacted more often.
- Results describe this NanoBench v0.7.1 task set and observed setup only.
File integrity
Use SHA256SUMS.txt to verify the two published Jinja files. release_manifest.json records the benchmark scope, current hashes, and provenance limits.
Model tree for jucabala457/flint-chat-templates
Base model
RecursiveSI/flint


