trymirai/Flint-S

Flint S

Mirai S quantization of Recursive Flint, a mixture-of-experts coding-agent model with 39.8B total parameters and approximately 2.5B active parameters per token.

The root-level package is Flint S 13.3: 13.157 GB of weights/shared tensors, 13.173 GB including tokenizer and metadata. It uses a custom packed S format and includes a standalone reference decoder.

Validation status: the KL and agent results below were measured on this checkpoint's dense dequantized evaluation build. The packed files passed reconstruction comparisons against that build, but decoded-package RomaKL and direct runtime integration remain unvalidated. These files are not a drop-in Transformers, GGUF, MLX or uzu checkpoint. The supplied decoder reconstructs tensors; it does not provide a serving engine.

KL / teacher fidelity

Lower KL is better. Top-1 is agreement with the teacher's next-token choice, not task accuracy.

Quantization	Size (GB)	On-policy RomaKL ↓	Teacher top-1 agreement	Dataset-text RomaKL ↓
Mirai Flint S 13.3, dense evaluation build	13.17	0.09117	91.49%	0.23781
Unsloth UD-IQ1_M	11.35	0.20225	86.72%	0.41156
Unsloth UD-IQ2_XXS	11.86	0.17714	87.51%	0.37993
bartowski IQ2_XXS	12.03	0.27529	84.46%	0.58081
bartowski IQ2_XS	12.74	0.22879	86.14%	0.52038
bartowski IQ2_S	13.39	0.21386	86.62%	0.48050
Unsloth UD-Q2_K_XL	13.56	0.13507	89.23%	0.29978
Unsloth UD-IQ3_XXS	14.54	0.12238	89.97%	0.33076
bartowski IQ2_M	14.88	0.15442	88.64%	0.38832
bartowski Q2_K	15.48	0.17558	87.84%	0.39941
bartowski IQ3_XXS	17.49	0.08920	91.42%	0.25920
bartowski IQ3_XS	18.32	0.06739	92.50%	0.25420
bartowski Q3_K_S	18.44	0.06659	92.57%	0.23567
Unsloth UD-Q3_K_XL	18.49	0.05566	93.30%	0.19121
bartowski Q3_K_M	19.53	0.06101	92.74%	0.21117
bartowski Q3_K_L	20.59	0.05018	93.46%	0.18979
bartowski IQ3_M	21.05	0.05041	93.51%	0.19371
bartowski IQ4_XS	22.45	0.02505	95.52%	0.13466
Recursive nvfp4	22.86	0.06065	93.09%	0.24904
Recursive int4-q4_1	25.99	0.05691	93.21%	0.22758
Q8_0	42.30	0.00479	97.92%	0.05890

The GGUF comparisons were measured through llama.cpp; Mirai S uses a dense evaluation twin. Recursive entries follow their recorded evaluation paths. These are cross-runtime measurements. Packed decoder KL is still pending. Both contracts use the same 100-prompt pool, rendered with Flint's chat template. The headline on-policy contract scores 76,946 tokens from Flint-generated continuations. The dataset-text contract scores 34,631 tokens from dataset responses.

Teacher cache and dense quantized evaluation use FP32 activations to reduce Flint's numerical instability on off-policy text. The BF16-activation teacher's measured floor against that cache is approximately 0.00359 KL / 98.36% top-1 on-policy and 0.0558 KL / 96.03% top-1 on dataset text. These numbers are specific to these contracts and precisions; they do not establish runtime-independent fidelity.

Local agent benchmark

These are Mirai's in-house agent/task evaluations, not official Terminal-Bench, SWE-bench or BFCL scores. The S column evaluates the dense dequantized twin of S3-13.3-qB, the checkpoint with 51/55 hard problems and 444/656 assistant-week v2 tasks. GGUFs use llama.cpp; the teacher uses SGLang. The packed decoder has not been agent-benchmarked.

Evaluation	Flint S 13.3	Unsloth UD-Q2_K_XL	bartowski IQ3_XXS	Q8_0	BF16 Flint
Function calling	92%	97%	94%	96%	91%
Function calling, hard	75%	76%	73%	75%	77%
Hard-problem battery	51/55	44/55	38/55	52/55	49/55
AppWorld, 80 tasks × 3	158/240	158/240	Unvalidated¹	164/240	187/240
Assistant week v2, native tools	444/656	242/656	333/656	358/656	460/656
Assistant week v1, text markers	41/136	18/136	36/136	57/136	60/136
Creative build, fully correct	4/4	4/4	4/4	4/4	4/4
Agentic build, declared done	4/4	2/4	3/4	4/4	4/4

¹ The bartowski AppWorld record completed 240 episodes in only 108 seconds. That result requires harness/protocol review and is excluded as a valid model-quality comparison. It is not presented as 0% task ability.

S and BF16 agent runs used SGLang; GGUF agent runs used llama.cpp. Backend, numerical and tool-protocol differences can affect results. These are recorded point estimates; this card does not claim statistically significant differences. Function-calling sets contain 175 standard and 150 hard cases, each spanning thinking/non-thinking runs and repeated samples (700/600 scored outputs). The hard battery contains 11 problems with five attempts each. Assistant v1 uses four seeds; v2 uses sixteen. AppWorld repeats each task three times. Creative/build sets contain only four tasks each. A declared-done result is not a render-quality pass; no latest-S agentic render result was graded. Historical text-marker sessions can loop or exhaust context, so v1 and native-tools v2 are reported separately.

Sampling uses temperature 1.0, top-p 0.95, top-k 20, with the original Flint template and task-specific token/step budgets. Aggregate scores, counts, checkpoint labels and hashes of the original result records are provided in benchmarks/results.json. Training used Flint-generated coding, reasoning and tool traces; these are local evaluations, not an independent external audit.

Files and format

• model-00.safetensors … model-49.safetensors: packed weight shards.
• shared.safetensors: codebooks and shared rotation parameters.
• flint_s.json and quantized-*.json: format and tensor metadata.
• flint_s_decode.py: standalone CPU reference decoder, requiring PyTorch and safetensors.
• Original configuration, tokenizer, chat template and license.
• PACKAGE_SHA256.json: file sizes and SHA-256 hashes for the latest payload, model card, decoder and benchmark summaries (excluding the manifest itself and legacy directories).

Trellis formats include V4 restart-64 T6/T7/T8 and connected V2 k3/k4, with Kronecker rotations and row scales; other quantized tensors use the package's int8 format. See the decoder's format documentation for the precise bit order and reconstruction arithmetic.

Packing reuses the exact previously recovered k3/k4 codes, and each of the 50 packed shards is decoded and compared against this checkpoint's dense twin. See PACKING_RESULT.json for the completed reconstruction checks. This establishes a weight-format check; decoded-model logits and direct runtime integration still require validation.

To inspect/reconstruct the latest package without downloading legacy dense exports:

hf download trymirai/Flint-S --local-dir flint-s \
  --include 'model-*.safetensors' shared.safetensors flint_s.json \
  flint_s_decode.py config.json 'tokenizer*' chat_template.jinja generation_config.json
python -m pip install torch safetensors
python flint-s/flint_s_decode.py flint-s flint-s-decoded

Reconstruction writes approximately 80 GB of dense tensors. The decoder output needs layout/configuration integration before serving; no one-command serving path for this packed layout is claimed.

License

Apache 2.0; see LICENSE. Base model: RecursiveSI/flint. This repository contains Mirai quantization artifacts and local evaluation results; credit for the base model belongs to Recursive.