Orbit · MMLA-Orbit-1B

Orbit is the first released MMLA model: a 1B-class model with authoritative memory, a learned policy-state initialization, and executable memory-manager adaptation. MMLA separates the information a model retains from the state that governs how it acts. This checkpoint brings those components together in one model package.

Runtime · MMLA report · LARC report · MMLA-RSI report

Architecture

Component	What it does
MMLA / PDSA	Separates policy state Φ and authoritative memory M, with explicit memory reads, bounded writes, and joint-action interfaces.
LARC / RTT	Provides a trained rank-4 residual initialization and problem-local feedback updates, with reset to the saved trained initialization.
RTMA / MMLA-RSI	Retains arrived experience in M and searches executable memory-management rules. An independently admitted successor can take over the management role.
Frame / Stream / Joint	Supports complete-frame and causal-stream input, CSBC segmentation, and combined numerical/memory actions.

MMLA is mixer-agnostic; Orbit uses a Transformer-based mixer with trained MMLA components. The package includes the task reader, trained proposer, original proposer control, independent qualification reader and shared input-memory readout. Its default deployment keeps the trained Φ fixed while using M. Numerical RTT and joint interfaces are available; predictive Q/J heads require qualification for their actual reader and task. Architecture support is separate from evidence of performance gains or successful recursive improvement.

Quick start

Download the complete repository, then install its pinned runtime:

python -m pip install -r /path/to/Orbit/requirements.txt

from mmla_orbit import load_orbit, load_manager

model = load_orbit("/path/to/Orbit", device="cuda")
manager = load_manager(model, "/path/to/Orbit/deployment.json")
state = model.new_task_state(seed=17, tenant="demo", rule=manager)

The host supplies task observations and real feedback through the task interfaces. See the RSI interface for bounded proposal, evaluation and rule succession. The current default task adapter uses ALFWorld-style observations; additional environments need an appropriate adapter.

Conventional language benchmarks

Model	ARC-E	ARC-C	HellaSwag	WinoGrande	MMLU
Orbit-1B	62.08	36.18	47.44	55.17	52.06
MiniCPM5-1B-SFT	61.11	35.75	47.77	56.12	53.01
MiniCPM5-1B (RL/OPD)	60.44	36.01	47.69	56.27	52.91
Qwen2.5-1.5B-Instruct	75.93	46.50	68.33	62.83	60.13
SmolLM2-1.7B-Instruct	62.92	43.86	71.77	68.35	48.40

Accuracy percentages on the complete 2,376 / 1,172 / 10,042 / 1,267 / 14,042 example panels. Evaluation uses zero-shot completion likelihood, empty M, fixed Φ, no chat template, and FP32 execution of the saved weights. ARC and HellaSwag use character-normalized likelihood; WinoGrande uses common-suffix likelihood; MMLU is micro accuracy over 57 subjects. These previously evaluated panels are regression comparisons. Qwen and SmolLM2 have more parameters. Exact counts, protocol and model identities are in benchmark results.

Agentic evaluation

Orbit uses fixed trained Φ, eight-slot authoritative M and the deployed RTMA manager. Eight external models use their saved chat/tool templates and ordinary conversation history. The table reports complete evaluations of these configured systems; their prompts, context limits, runtimes and hardware differ.

Model	ALFWorld success	BFCL success	ScienceWorld mean	ScienceWorld success
Orbit-1B (M + RTMA)	41.79% (56/134)	0.00% (0/200)	-28.28	1/90
MiniCPM5-1B-SFT	0.00% (0/134)	1.50% (3/200)	-16.50	0/90
MiniCPM5-1B	0.00% (0/134)	2.00% (4/200)	-8.28	0/90
Qwen2.5-1.5B-Instruct	0.00% (0/134)	0.50% (1/200)	-25.36	1/90
SmolLM2-1.7B-Instruct	0.00% (0/134)	0.00% (0/200)	1.92	0/90
LFM2.5-1.2B-Instruct	0.00% (0/134)	1.50% (3/200)	2.92	1/90
SmolLM3-3B	0.00% (0/134)	3.50% (7/200)	-22.06	2/90
Qwen3.5-4B	3.73% (5/134)	46.50% (93/200)	-9.31	3/90
Qwen3.5-9B	24.63% (33/134)	60.50% (121/200)	-24.30	8/90

ALFWorld is a 134-task seen regression panel. BFCL covers 200 multi_turn_base cases. ScienceWorld covers 30 task classes with three fixed test variations each, scored by the native terminal score (−100 to 100); task successes are shown separately. These panels have been reused during diagnosis. They measure the declared bounded protocols, with no claim of unseen generalization, a full official leaderboard or architecture-only advantage.

ALF protocol correction. Orbit now uses the original action instruction defined in its packaged ALFWorldFormatter. A complete same-weight comparison found 56/134 successes with that instruction and 35/134 with the evaluation-specific rewrite: 30 tasks succeeded in both arms, 26 only with the original and 5 only with the rewrite. The table follows the packaged interface; both complete arms are retained. External ALF results retain their revised instruction, so this column is not a prompt-matched model comparison. Historical Orbit counts 54/35/36 belong to their original runs; the new comparison does not reconstruct all historical physical states.

Tool and science failures remain visible. Orbit BFCL is still 0/200. Complete schema-normalization and pre-calibration-readout comparisons did not restore task success. The ScienceWorld table uses the latest complete revised-instruction arm: −28.28 mean and 1/90 success. The original-instruction arm had −6.41 mean but 0/90 success, mostly avoiding negative terminations; it is retained as a diagnostic. The earlier revised-instruction run had the same mean but 0/90 success and seven different per-task scores. These are separate runs, not values selected per task.

All tasks allow at most 64 replies; ALFWorld/ScienceWorld allow 64 generated tokens per reply and BFCL 512. Orbit contexts are 4096/8192/16384 tokens for ALFWorld/ScienceWorld/BFCL; SmolLM2 uses 8192 and the other external models 32768. BFCL parses saved text replies and returns tool feedback as user observations: [] or a parsing rejection ends a user turn, and the 21st nonempty reply triggers the per-turn stop. The exact native tool-call/tool-result messaging protocol is not reproduced. MiniCPM5-1B uses the saved SFT template override. ScienceWorld initializes from the variation ID; the host seed argument is not forwarded, and exact historical JVM-state replay is not established.

The 3392 external endpoints are retained in full, including the independently checked Qwen3.5 BFCL correction to 93/200 and 121/200. That correction permits a prose prelude before a complete native function-call block; the new runs used a local RTX4090 while historical Qwen runs used RTX PRO6000. Nine SmolLM3 BFCL tasks used explicit continuation; unreturned-request compute is incompletely known. Source identities, all 3816 per-task outcomes, limits, precision, continuation records and costs are in the result files. Exact runtime snapshots and raw trajectories remain in the private research archive; the public indices support score recomputation but are not a standalone end-to-end harness release.

No new weights or RSI successors are installed for these results. Numerical RTT and recursive adaptation remain separate evaluation questions.

Scores, protocols and performance · Per-task results · Both instruction arms · Historical v3 · Earlier protocols

Files and license

model.safetensors contains all model tensors, including omega, qualification and the internal roles. config.json records the architecture, role layout and exact runtime version. Tokenizer and generation files are at the repository root. Runtime code is maintained in MMLA-RSI; use load_orbit to restore the complete model and its memory interfaces.

The consolidated archive preserves the original tensor values and role identities. It does not require separate adapter weight files. deployment.json supplies the packaged memory-management rule, and MANIFEST.json records file hashes. Optimizer states, raw task traces and internal reports are outside this inference package.

MMLA additions are licensed under AGPL-3.0. Upstream components retain their licenses and attribution in NOTICE. Please cite the three reports linked above; CITATION.bib provides BibTeX.