netprojects/h2o-lightning-4b-mlx-q4

H2O-Lightning-4B — MLX 4-bit

A 4-bit MLX conversion (affine, group size 64; mlx-vlm 0.7.8) of h2oai/h2o-lightning-4b v1.4.1 as updated by H2O on 28 October 2026 (commit d7f1b05aa0b4e789b3896116f309ef842440e260; the weights are those of the original v1.4.1), for Apple Silicon. 2.9 GB. The smallest build, for one-way reading. Box questions work.

Published by NetProjects (Internet Projects Ltd). Not affiliated with H2O.ai.

This is a decision model, not a chat model. It returns a probability for each answer option (or, for a Box question, a bounding box), not free text.

Serve it with the H2O shim, not with mlx_lm.server or a plain mlx_vlm generate call. The generic "Use this model" snippets Hugging Face shows for MLX models do not return the label probabilities this model is made for. The commands below set it up with H2O's official shim, unmodified, on top of h2o-lightning-mlx, which reads the label log-probabilities with a float32 head, as the original does on vLLM.

JevBench public (231, JevBench v1.2.1, --adapter typesafe), one-way
this checkpoint	205 / 231 (48/48 · 71/72 · 86/111)
original weights on MLX (bf16)	201 / 231
H2O's model card (vLLM), one-way	202 / 231
vs bf16 on MLX	221 / 231 answers identical; probability drift mean 0.043, p95 0.176

Which build? This build serves one-way reading: start it with TWO_WAY=0. For H2O's two-way reading (205 / 231; H2O's own result 206) and for the closest match to the original's calibration, use the 8-bit build (4.8 GB).

Box questions (always on since v1.4.0) work on this build too: on H2O's receipt example it finds the total-amount line with confidence 0.90 (8-bit: 0.97).

Speed (MacBook Pro M5 Max, 27 October 2026, through serve.sh, one request at a time): about 59 ms per JevBench item, a 2,000-token record in 0.36 s, a 32K-token state in 7.4 s. MLX peak memory 5.7 GB at 4K tokens and 6.8 GB at 32K.

Images. Image input works through the unchanged shim. On H2O's own receipt example this build gives the same answers as the original weights. It is not a reproduction of H2O's image benchmark.

About v1.4.1 (from H2O's card): the v1.4.0 fine-tune continued on more training data, with unchanged calibration constants; stronger on retrieval and on six task families the added data covers, level on H2O's held-out decision skills overall. This build ships the training tokenizer (tokenizer/), which the shim checks at start.

28 October 2026 revision. H2O updated v1.4.1 in place; the weights did not change, so this build was not reconverted, and it now ships the updated shim, serve_config.json and the new selector_v2.py. Over-long records are now trimmed to fit the context by BM25 selection instead of a fixed cut at 32,000 tokens, so a request never fails for length: on a 60,000-token test record with one key line in the middle, this build now answers correctly (0.95, one-way). Image answers are read at temperature 1.0 instead of 0.65: the top answers stay the same and the confidence values are lower (receipt item count 0.94 → 0.81 one-way). Text requests that fit the context are read exactly as before.

Limits (as for the original): mostly evaluated in English; works best with about 16 answer options or fewer; no explanations and no video input.

git clone https://github.com/internet-projects/h2o-lightning-mlx && cd h2o-lightning-mlx && ./setup.sh
hf download netprojects/h2o-lightning-4b-mlx-q4 --local-dir mlx-q4
TWO_WAY=0 MODEL=mlx-q4 ./serve.sh # http://127.0.0.1:8741/v1/systemone

Versions. main is H2O v1.4.1 with H2O's 28 October revision (tag v1.4.1); the original 27 October v1.4.1 files stay available with --revision v1.4.1-initial. Earlier builds stay available: H2O v1.4.0 (one-way, 202 / 231) with --revision v1.4.0, H2O v1.3.1 (one-way, 199 / 231) with --revision v1.3.1, and H2O v1.2.3 (207 / 231) with --revision v1.2.3; use each with the matching tag in the GitHub repo.

Credits: H2O.ai (H2O-Lightning-4B, Apache-2.0), Qwen team (Qwen3.5-4B, Apache-2.0). This conversion is not affiliated with H2O.ai.