Flint Terse Minimalist (MLX)
Recursive Flint, fine-tuned to think briefly and in shorthand inside <think> while keeping its final answers. This is a complete MLX model for Apple Silicon: download it in LM Studio or load it with mlx-lm, with nothing else to install. The GGUF version is gyohng/flint-terse-minimalist.
"Hi!" base: 123 thinking tokens ("The user has just said "Hi!" — a simple greeting. ...")
terse: 15 thinking tokens User greets; respond friendly; ask how can help; brief; casual.
flint-terse-minimalist thinks the least: about 80–90% less in every category, at about 4 code tasks below the base out of 110. For full base-level coding accuracy, see flint-terse.
Results
Measured on the int4 base with the adapter in llama.cpp. Scores are the mean of three seeds with one sample per task (temperature 1, top-p 0.95, top-k 20). The thinking reduction is relative to the base on the same tasks.
| Code (110 tasks) | Math (120 tasks) | Less thinking: code | math | prose & chat | |
|---|---|---|---|---|---|
| Flint int4 (base) | 102.3 | 103.3 | — | — | — |
| flint-terse | 102.7 | 109.7 | 62% | 79% | 85% |
| flint-terse-minimalist | 98.0 | 108.0 | 81% | 86% | 85% |
flint-terse keeps the base's coding accuracy. flint-terse-minimalist thinks the least, at about 4 code tasks lower. Final replies stay complete, and no <think> tags leak into them.
What this model is
The MLX int4 base RecursiveSI/flint-int4 with the LoRA merged in (30.6 GB). The merge rewrote attention q/k/v/o, the shared expert, the two dense MLPs and the routed experts' down-projections in the last 16 MoE layers. They are stored unquantized as float16: re-quantizing them to 4 bits, as mlx_lm.fuse does, would round away most of the fine-tune, and bfloat16 would round away 5–18% of it. Every other tensor is the base's original int4. As with any mlx-lm LoRA, activations then run in float32. The merge was verified against llama.cpp running the GGUF adapter on the int4 GGUF: mean |Δlogp| 0.010 and 100% top-token agreement. The chat template is Flint's, minus a replace("\x00", "") that LM Studio's template engine does not support.
The unmerged adapters (GGUF for llama.cpp, MLX for mlx-lm --adapter-path) are at gyohng/flint-terse-minimalist-lora.
Usage
LM Studio: search for gyohng/flint-terse-minimalist-mlx, download, load. Set temperature 1, top-p 0.95 and top-k 20 (the base model's defaults).
mlx-lm:
mlx_lm.generate --model gyohng/flint-terse-minimalist-mlx --temp 1 --top-p 0.95 --top-k 20 --max-tokens 4096 \
--prompt "Write a function that merges overlapping intervals."
mlx_lm.server --model gyohng/flint-terse-minimalist-mlx # OpenAI-compatible server
How it was made
Training ran directly on the int4 coefficients with rank-32 FP32 LoRA on attention, the shared expert and the dense MLPs, plus rank-8 LoRA on the routed experts' down-projections in the last 16 MoE layers. The model learned from its own answers by reinforcement learning: an answer earned reward only if it passed its tests and fitted a shrinking thinking budget. A short supervised pass then taught terse thinking for prose and chat. Code, data and a one-script reproduction are at https://github.com/gyohng/flint-terse.
License
Apache-2.0, the same licence as Flint. This fine-tune adds no further conditions. Flint is by Recursive.