ChengyiX/coder390-fable-m4max-agent-study

Fast on the Card, Slow in the Agent

A Vision and Speculation Study of Coder390 and TURBO-Fable on Apple Silicon

Chengyi Xu and Nexora — Nexora Research, October 2026

An AI-assisted experimental deployment report comparing two requested Qwen3.8-27B replacements against an installed normal/abliterated Ornith-1.5-35B-A3B pair on Apple M4 Max, 40-core GPU, 64 GiB unified memory.

Public hosting is not peer review or journal acceptance. No DOI is claimed. No model weights are redistributed.

Download the paper

• Paper PDF
• LaTeX source, using nexora-paper.sty with the same layout as earlier reports and updated Nexora branding
• Numerical results
• Pinned model/runtime provenance
• Vision probe results

Main result: do not replace the working deployment yet

Both requested candidates loaded with thinking and vision; all seven small llama.cpp vision probes answered a known diagram correctly. That did not establish coding-agent usability.

Model/configuration	Whole task	Final visible stream	Full final-cycle visible rate	Repair result
Installed normal Ornith, thinking low	218.08 s	30.57 tok/s	23.86 tok/s	Independent gate + 275 tests passed, 1 skipped
Installed abliterated Ornith, thinking low	204.68 s	32.33 tok/s	28.38 tok/s	Independent gate + 275 tests passed, 1 skipped
Coder390, five tested configurations	~900 s each	Unknown: no completed final answer	Unknown	Five timeouts; injected defect remained
Fable, b11036 MTP4	889.64 s	4.32 tok/s	2.44 tok/s	Independent gate + 274 tests passed, 1 skipped
Fable, b11538 MTP2	892.05 s	8.38 tok/s	6.22 tok/s	Independent gate + 274 tests passed, 1 skipped
Fable, b11538 no speculation	795.80 s	6.50 tok/s	3.15 tok/s	Independent gate + 274 tests passed, 1 skipped

The task was a real Pi repository repair, routed through Magpie with normal extensions and tools, in a fresh isolated clone for every run. Initial requests were approximately 25k tokens, not blank prompts. A deliberately injected regression failed before the task; the parent checked the repaired source, independent gate, and full suite afterward.

These are configuration-level descriptive results, not a causal ranking of model intelligence. There is one trial per configuration, different sampling and engines, and variable workstation load. Candidate cold first-turn waits of roughly 365–579 seconds dominated the deadline. Coder390 may complete with different runtime/settings or more time; these failures are not proof that it cannot solve the task.

Important compatibility findings

• The publishers' headline speed measurements used NVIDIA RTX PRO 6000 / RTX 5090, not this Mac.
• Coder390's static FP8 headline scores differ from its Q4 GGUF table. We did not independently rerun GPQA/MMLU/LCB.
• Author recipe -c 409600 -np 4 means 102400 tokens per slot, not 409k per individual session. Native config declares262144 positions.
• SGLang does support Apple MLX, but its inspected in-tree DFLASH handler excludes MPS. No unsupported SGLang+DFlash2 result is invented.
• Native Splash admits only about208889 context tokens for Coder390 at28GiB;32GiB admits262144 startup. That is allocation readiness, not actual256k qualification.
• The tested Coder390 native DFlash2 path was text-only, using a compatible BF16 draft, not the author's FP8 draft or requested Q8 projector. Vision proof from llama.cpp does not transfer to it.
• The existing native importer rejected Fable's BF16 output head. No hidden requantization or unverified kernel patch was used.
• MTP acceptance is not whole-task speed. In Fable's measured newer-build runs, no-MTP completed the whole task sooner than MTP2 despite lower final-stream speed.

Reproducibility and privacy

See REPRODUCIBILITY.md, RESULTS.md, and protocol.md.

The public ledger is a whitelist-only export. Private repository source, raw source-bearing transcripts, tool arguments/results, secrets and local user paths are withheld. The original private task cannot be reproduced byte-for-byte publicly. Using your own authorized repository is a new method replication, not identical evidence.

No failed run is dropped. null means unknown, not zero. Visible-stream rates exclude preceding reasoning/prefill; full final-cycle rates include waiting and reasoning. Neither is whole-task visible throughput. Tool-only parser bursts are not decode evidence.

Selected method tests:

cd code
python3 -m unittest test_export test_pi_measure test_public_evidence
python3 make_figures.py

The tests validate export/privacy and measurement logic, not model inference. Figures and the LaTeX numerical table are regenerated from all ten public ledger rows.

Published checkpoints evaluated

• nerkyor Coder390 EfficientThink
• DavidAU TURBO-Fable Cold-Fusion
• Normal Ornith GGUF
• Separate abliterated Ornith GGUF

Full revisions, hashes, actual formats and publisher-evidence scope are in the provenance ledger. Terms such as "distilled," "uncensored," and "Heretic" describe publisher lineage, not proprietary-model identity or a guarantee of particular behavior. No harmful-generation/refusal benchmark was performed.

Outcome and disclosures

Neither candidate was promoted or permanently registered in Magpie. Temporary test providers were removed and the prior service restored. Old Ornith weights were preserved. No cloud defaults or image workflow were changed. No monitors or delegated agents were used.

AI tools materially assisted research, code, experiments, source comparison, analysis and drafting. No independent human manuscript review, funding/conflict attestation, contributor-role declaration or peer acceptance is asserted. Attribution follows the project owner's requested authorship.

Original code is MIT; original manuscript and sanitized numerical evidence are CC BY4.0. See LICENSE for mixed-artifact terms. Referenced models/software retain their own licenses. Use CITATION.cff to cite this artifact and the specific published revision.