Algocean-Qwen3.6-35B-A3B-mlx-OptiQ

Quality-first OptiQ MLX release of the Algocean fine-tune derived from Qwen/Qwen3.6-35B-A3B.

The source model for this quantized release is aisamdasu/Algocean-Qwen3.6-35B-A3B-mlx, which was produced by merging the best LoRA checkpoint from the Algocean SFT run. This artifact keeps the same text-focused MLX scope as the merged MLX release: the current mlx-lm Qwen3.5-MoE conversion stores language-model weights and omits model.visual.

Use

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("aisamdasu/Algocean-Qwen3.6-35B-A3B-mlx-OptiQ")

messages = [{"role": "user", "content": "Create a concise LangGraph plan for a repo refactor."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, max_tokens=512, verbose=True)

Artifact

item	value
Format	MLX safetensors
Quantization	OptiQ mixed precision
Shards	7
Weight size in MLX index	36,036,335,872 bytes
Parameters in MLX index	34,660,608,768
Converter output size	34,367.2 MB
Visual weights	omitted
Integrity file	SHA256SUMS

OptiQ Settings

key	value
Method	optiq_mixed_precision
Source artifact	aisamdasu/Algocean-Qwen3.6-35B-A3B-mlx
Target BPW	4.5
Achieved BPW	4.733167
Candidate bits	4, 8
Allocation	297 tensors at 4-bit, 94 tensors at 8-bit
Reference	uniform_4bit
Group size	64
Calibration source	aisamdasu/algocean-fable5-traces train split
Calibration mix	512 train records sampled before eval separation
Exact calibration	24 samples, 48 calibration sequences, seq_len 512
Evaluated tensors	391
Skipped sensitivity entries	120 entries without BF16 source match

This is the quality-first OptiQ export from this run. Instead of publishing a smaller static or uniform 4-bit artifact, the exact OptiQ pass measured layer sensitivity on the FABLE/LangGraph calibration mix and kept the more sensitive tensors at 8-bit. The result is larger than a uniform 4-bit model, but preserves more capacity in the layers that measured as most sensitive for this calibration workload.

Training

key	value
Base model	Qwen/Qwen3.6-35B-A3B
Dataset	aisamdasu/algocean-fable5-traces
Train / eval rows	30,265 / 512
Selected source rows	Crownelius 26,938 + kelexine 3,839
Dedup skipped	1,721
Fine-tune method	LoRA SFT
LoRA target modules	q_proj, v_proj
LoRA rank / alpha / dropout	16 / 32 / 0.0
Max sequence length	16,384
Optimizer / schedule	AdamW fused / cosine
Learning rate / warmup	1e-5 / 0.03
Batch / grad accumulation	1 / 1
Max steps	4,000
Eval / save interval	200 / 200
GPU	Modal B200 x1
Runtime	7,934.71 sec

Checkpoint Selection

The published weights use checkpoint 3600, not the final step 4000.

step	eval loss	eval PPL	note
200	1.300712	3.671911	first eval
3200	1.067631	2.908480	near plateau
3400	1.067996	2.909543	regression
3600	1.066911	2.906388	selected best
3800	1.067495	2.908084	worse than best
4000	1.067005	2.906661	final, slightly worse

Best checkpoint rationale:

• 3600 has the lowest validation loss: 1.066911.
• Final step 4000 is very close but still higher: +0.000094 eval loss.
• The curve flattened after 3200, so selecting 3600 keeps the best validation point without chasing extra train loss.

Metrics

metric	value
Best eval loss	1.066911
Best eval perplexity	2.906388
Final eval loss	1.067005
Final eval perplexity	2.906661
Trainer average train loss	0.494604
Last logged train loss	0.996137
Train samples/sec	0.504
Train steps/sec	0.504
Eval samples/sec at final eval	2.088

Comparison Notes

This OptiQ artifact is selected for fidelity under the available local export options: best validation checkpoint first, then exact OptiQ sensitivity allocation instead of static quantization. It is not claiming a public benchmark win over the base model, the merged MLX model, or the upstream OptiQ model; that would require the same held-out benchmark harness across all models. The comparison recorded here is based on the completed training run's validation curve and the OptiQ allocation metadata.

Intended Use

This model is intended for local MLX inference experiments around agentic coding, tool-use traces, and LangGraph-style planning. It is not a general safety-tuned assistant release.