Flint-40B-A2.5B-abliterated REAP-224 (Rec 4-bit)

dima6312/Flint-40B-A2.5B-abliterated-Rec-4bit with 32 of the 256 routed experts removed in every MoE layer (256 → 224), chosen by REAP (router-weighted expert activation pruning) on coding and coding-agent text. The layout is unchanged (affine int4, group size 64, the same as RecursiveSI's int4 release); only the expert tensors and router rows of the 47 MoE layers are sliced.

Code and full write-up: github.com/RINNECODER/flint-research: pruning scripts, masks, routing stats, eval logs, and the Locals patches for DFlash 2.

It is about 12% smaller (19 GB instead of 21 GB). When experts are streamed from SSD it decodes faster, and it scored the same as the 256-expert model on a small HumanEval+ test.

Results

All measurements are on an M5 Pro Mac with 24 GB, using Locals with SSD expert streaming (a 6 GB expert cache) and DFlash 2 speculative decoding where marked.

	256 experts (original)	REAP-224 (this repo)
HumanEval+ pass@1, 40 tasks (every 4th), greedy, full reasoning	35/40 (87.5%)	35/40 (87.5%)
Decode tok/s, DFlash 2 off	25.4 / 29.3	30.1 / 30.4
Decode tok/s, DFlash 2 on	30.1 / 34.5	36.8 / 36.8
DFlash 2 draft acceptance	0.64	0.68
KL from the 256 model on held-out code and agent text	—	0.92
Top-1 agreement with the 256 model	—	78.0%

How the tok/s numbers were taken: 2 coding prompts × 2 runs, 1,024 tokens each. Where two numbers are shown, they come from separate sessions.
KL baseline: the KL figure is not near zero. Even changing only the chunked-prefill numerics of the 256 model gives a KL of 0.42 against itself.
Correction: re-scoring the saved evalplus results on only the 40 sampled tasks gives 36/40 for the 256-expert model, not 35/40. The table above predates that re-count. REAP-224 stays at 35/40; the one task it lost is HumanEval/48. Details are in the repo's docs/eval.md.
Small eval: with only 40 tasks this is a sanity check, not proof that quality is unchanged. Long agentic sessions were not tested.

Deeper cuts were tried and dropped:

Experts kept	HumanEval+ (n=40)	Decode tok/s with DFlash 2
192	32/40	36.5
160	29/40	38.3

Applying the published RESMP RCO-224 mask to this model kept a different set of experts. On coding text it did no better than dropping 32 experts at random:

Mask	KL from the 256 model	Top-1 agreement
RESMP RCO-224	1.49	68.3%
Random 224	1.47	68.6%
REAP-224 (this repo)	0.92	78.0%

How the experts were chosen

Calibration data: 217,606 tokens taken from codeparrot/codeparrot-clean-valid, nebius/SWE-agent-trajectories and SWE-Gym/OpenHands-SFT-Trajectories. Evaluation used different rows of the same datasets.
Scoring: for each expert, the mean over routed tokens of gate weight × the L2 norm of the expert's output. The 224 highest-scoring experts in each layer were kept.
What was removed: the experts dropped carried 10.1% of routing slots on the calibration text (a random 32 carry 12.6%). REAP keeps some rarely used experts that matter when they fire, so it is not the same as dropping the least-used experts (that set carries 3.5% but tracks the original model less closely: KL 0.98 vs 0.92).
Expert ids: keep_ids.json maps the new expert indices to the original ones. keep_ids[layer - 2][j] is the original id of new expert j; layers 0–1 are dense.

Coding was the target, so other domains (non-English text, security knowledge, maths) may lose more than the coding numbers above show. Refusal behaviour comes from the abliterated base and was not re-tested after pruning.