Metaflora Incubus v1

Metaflora Incubus v1 is a compact local model system built for code, structured tool calls, agentic workflows, and Russian-English generation. Its routed candidate was post-trained on 6,750 benchmark-disjoint records spanning bilingual instruction, rewriting, executable code, native tool dialogues, agentic replay, and refusal reduction.

Retained 48-case diagnostics report 100.00 agentic, 95.83 code, 95.00 tool use, 87.50 English, 81.25 Russian, and 81.25 text quality. The Q5_K_M core serves through a local OpenAI-compatible endpoint, works offline after installation, and requires no mandatory cloud moderation service or remote policy gateway. Voice and vision remain optional downloads, so the primary text package stays at 3.075 GB.

Highlights

• Compact 4B Q5_K_M deployment.
• Deterministic structured tool use and agentic search evaluation.
• Local operation without a mandatory cloud moderation service or remote policy gateway.
• OpenAI-compatible serving through a recent llama-server.
• Optional signed voice and vision packages.
• Artifact-bound benchmark receipts, checksums, and immutable revisions.

Model overview

Property	Value
Type	Causal language model
Parameters	4B
Release format	GGUF
Quantization	Q5_K_M
File size	3.075 GB
Recommended memory	16 GiB
Validated context	8,192 tokens
Release languages	Russian and English

Training and data

The routed release candidate was produced through several post-training stages covering bilingual instruction following, text correction, executable code, structured tool calls, agentic workflows, and refusal reduction. Training records were selected separately from the fixed 48-case evaluation bank.

Training scale at a glance

Recorded quantity	Value
Main continuation records	6,750
Language and writing records	4,000
Execution-oriented code records	1,500
Native tool and agentic records	1,250
Capability groups in the main mixture	6
Optimizer updates	1,589
Effective sequence slots processed	3,178
Maximum scheduled token positions	2,440,704
Tool-specialization records selected	1,090
Tool-specialization packed sequences	992
Fixed diagnostic cases	48
Cases per evaluated capability	8

The 2,440,704 figure is the schedule ceiling calculated as 1,589 updates × 2 sequences × 768 positions. It describes the maximum number of packed token positions presented by the schedule, not the number of unique natural-language tokens in the source corpus.

Main continuation mixture

Data group	Records	Share	Purpose
Russian instruction	1,500	22.22%	Russian requests, concise answers, and format control
English general instruction	1,500	22.22%	General instruction following and answer structure
Rewrite and correction	1,000	14.81%	Editing, grammar correction, compression, and clarity
Execution-oriented code	1,500	22.22%	Implementation, repair, tests, and exact output behavior
Native structured tool calls	750	11.11%	Tool selection and schema-valid arguments
Agentic replay	500	7.41%	Multi-step planning, observation handling, and completion
Total	6,750	100.00%

Grouped another way, the mixture contains 3,000 direct bilingual instruction records, 4,000 language and writing records after adding rewrite/correction, 1,500 code records, and 1,250 native tool plus agentic records. Language data therefore accounts for 59.26% of the main continuation; specialist code, tools, and agentic behavior account for the remaining 40.74%.

The complete mixture is content-addressed by SHA-256: 9141185743bd8681493caac268cfb0339e95b0fd1788179d60d37a5c5671b183. The training manifest fixes the category counts, ordering inputs, packing rules, and output destination so the run cannot silently switch datasets.

Main training run

Setting	Value
Sequence length	768 tokens
Effective batch	2 sequences per optimizer update
Optimizer steps	1,589
Warmup steps	47
Warmup share	2.96% of optimizer steps
Effective sequence slots	3,178
Maximum scheduled token positions	2,440,704
Peak learning rate	3e-5
Final training weights SHA-256	b28a1744927aa2d90f947d64163a1106136db4c18e1e94784d8adcd7639a3649
General-route package SHA-256	0915944be328c0ebef2087a6007cfe4f046c5b83709d5929f2504da51c15fd15

The completed run retained a receipt binding the dataset, configuration, final weights, and exported package. A runtime smoke test then loaded the exact package used by the general route before the 48-case diagnostic.

Tool specialization

The structured-tool continuation was selected from the pinned 6,750-record source rather than assembled from benchmark prompts:

Tool continuation group	Records	Share
Native tool-call dialogues	600	55.05%
Agentic replay	240	22.02%
English replay	150	13.76%
Russian replay	100	9.17%
Total	1,090	100.00%

These records were packed into 992 sequences of at most 512 tokens. The selection SHA-256 is 2a9ecf6254ba2bb2da46fbda7aebf53d211157a0a1a04c57dd5814236ac611b7. General English, Russian, and agentic replay remain in the mixture to reduce specialist overfitting. Tool output is checked against declared names and JSON schemas, with one bounded repair attempt before failure.

Native tool dialogues and agentic replay make up 77.07% of this specialist selection. The remaining 22.93% is English and Russian replay retained as a language-regression buffer. Packing converts 1,090 records into 992 sequences, with a maximum packed capacity of 507,904 token positions at 512 positions per sequence.

Capability isolation

The final system keeps ordinary language, code, and tool behavior on explicit routes. The general route handles Russian, English, text quality, and ordinary dialogue. The code route is selected only for programming requests because it performs better on code while reducing prose quality. The tools route is available only when the client supplies a declared tool schema.

This isolation prevents specialist tuning from becoming the default response policy. Every routed response can expose the selected route and artifact hash, and a requested specialist route fails closed if its package is missing.

Refusal-reduction stage

A dedicated refusal-reduction stage ran before the final multi-capability continuation. It targeted unjustified refusals on lawful, benign, and open-topic requests while preserving separate client-side permission checks for tools and external actions.

The interrupted training session did not retain the separate SFT-only versus refusal-reduction ablation receipt or the expanded open-topic report. The card therefore reports the retained 48-case refusal measurement and does not claim that every possible prompt receives an answer.

Evaluation and promotion controls

• The held-out diagnostic uses 48 fixed cases, eight per capability.
• Generation is fixed at temperature 0, seed 4242, and a 512-token output limit.
• Scores are recomputed from raw responses instead of copied from a summary.
• Missing cases, duplicate cases, truncated answers, forbidden terms, and artifact mismatches invalidate the affected evidence.
• A specialist package is not allowed to replace the general route without a full language and agentic regression check.

The training receipts support the routed candidate described in the benchmark section. The standalone public GGUF remains a separately hashed artifact until the routed packages are merged, signed, and published as one release.

Install

One-line installation

macOS on Apple Silicon:

curl -fsSL https://huggingface.co/metaflora/incubus/resolve/main/install.sh | sh

Text, voice, and vision:

curl -fsSL https://huggingface.co/metaflora/incubus/resolve/main/install.sh | sh -s -- --with-voice --with-vision

Enable vision on an existing installation without replacing the text weights:

incubusctl update --with-vision

The bootstrap is pinned to an immutable installer archive and verifies its SHA-256 checksum before execution. Linux, Intel macOS, and Windows automatic installation will be published only after signed runtime artifacts pass the same release checks.

OpenCode / open-source IDEs

The installer starts a loopback-only OpenAI-compatible service on the local machine and adds Incubus to the current OpenCode configuration. Restart OpenCode after installation, then select metaflora-incubus/metaflora-incubus-v1. An installation made with --with-vision also declares image input to OpenCode and starts the same local endpoint with the signed native mmproj-F16 vision projector. Attach PNG or JPEG files through OpenCode; text-only installations do not advertise image support.

Field	Value
Provider ID	metaflora-incubus
Display name	Metaflora Incubus v1
Base URL	http://127.0.0.1:18991/v1
API key	leave empty, or use local if the client requires a value
Model ID	metaflora-incubus-v1
Model display name	Metaflora Incubus v1
Extra headers	none

If OpenCode was installed after Incubus, register the provider without downloading the model again:

incubusctl integrate opencode

Check the service before connecting an IDE:

incubusctl status
curl http://127.0.0.1:18991/v1/models

Run a one-shot prompt directly from the terminal:

incubusctl run "Explain this repository"

Use incubusctl start, incubusctl stop, and incubusctl logs to manage the background service.

The same Base URL and Model ID work with open-source IDEs and local agent clients that accept a custom OpenAI-compatible endpoint, including Continue, Cline, Roo Code, Zed, Aider, and Open WebUI. Keep the service on 127.0.0.1; it has no API authentication because it is intended for local single-user use.

Opening links is a client-side agent capability rather than hidden network access inside the model. In OpenCode, enable its browser or web-fetch tool; the client retrieves the page and supplies its contents to Incubus. The local runtime never fetches arbitrary URLs on its own.

Quickstart

Ollama

Download metaflora-incubus-v1.gguf and create a Modelfile:

FROM ./metaflora-incubus-v1.gguf
PARAMETER num_ctx 8192

ollama create metaflora-incubus:v1 -f Modelfile
ollama run metaflora-incubus:v1

llama.cpp

llama-server \
  -m ./metaflora-incubus-v1.gguf \
  --host 127.0.0.1 \
  --port 11435 \
  -c 8192

The server exposes an OpenAI-compatible API at http://127.0.0.1:11435/v1.

Manually started llama.cpp clients

Field	Value
Base URL	http://127.0.0.1:11435/v1
API key	empty, or local if the client requires a value
Model ID	metaflora-incubus-v1

Keep the endpoint bound to 127.0.0.1. Public network exposure needs authentication, request limits, and a separate security review.

Agentic and tool usage

Agentic use requires a tool-enabled client. The model emits structured tool calls; the client retains control over filesystem, browser, shell, and network permissions.

Suitable workflows include:

• structured tool selection and JSON arguments;
• search tasks where a client supplies a search tool and returns sources;
• code generation, explanation, and review;
• local Russian and English drafting;
• private single-user inference.

The model itself has no network access. Every external action comes from the client and its configured tools.

Benchmark Results

Routed release candidate

The latest 48-case diagnostics use a frozen general route plus isolated code and tool routes. Each request is scored on the route intended for that capability. The results are bound to the following artifacts:

• General route artifact: 0915944be328c0ebef2087a6007cfe4f046c5b83709d5929f2504da51c15fd15
• Code and tool route artifact: d066766d4bbd0346b8a4a02a35d752e39ff7f70b064b71272d6fa43bdb9be526

Capability	Route	Score (/100)
Agentic workflows	General	100.00
Code	Code	95.83
Tool use	Tools	95.00
English	General	87.50
Russian	General	81.25
Text quality	General	81.25

These are direct diagnostics from the completed candidate runs. The code score comes from the isolated code route; enabling that route for ordinary prose would reduce English and text quality. The complete routed regression run and production signature are still pending, so these values are not presented as scores for the standalone GGUF download.

Standalone public GGUF verification

The currently downloadable Q5_K_M file has SHA-256 df850ca6f8d47b2d92db99fb623a36e9d35f3ad7737a588e7ff562ac5229b3fc. Its older signed receipt remains available under reports/ for artifact verification. It must not be mixed with the routed candidate scores above.

Evaluation setup

The deterministic suite contains 48 fixed cases, with eight cases per capability. Generation uses temperature 0, seed 4242, a 512-token output limit, and the declared runtime.

The final capability score is:

100 × sum(case scores) / 8

Scoring

• Code, English, Russian, and text-quality cases list required answer elements. Each score is the fraction of required elements found in the normalized response.
• Tool and agentic cases award 60% for the exact tool name and 40% for exact JSON arguments.
• Refusals, forbidden terms, truncated responses, duplicate cases, missing cases, and mismatched artifact receipts invalidate or zero the affected evidence.
• Aggregates are recomputed from raw model responses.

These results describe the Incubus release suite. They are not directly equivalent to MMLU-Pro, GPQA-Diamond, LiveCodeBench, BFCL, SWE-bench, Terminal-Bench, or TAU2.

The standalone GGUF has signed evidence under reports/. The routed candidate measurements are diagnostic until the production signing run is complete.

Open local operation and refusal behavior

The local release has no mandatory cloud moderation service, remote policy API, hidden server-side prompt filter, or vendor-controlled allowlist. Prompts and generated text remain on the user's machine unless the selected client invokes an external service. The model can operate offline after installation.

The earlier signed standalone run observed a 0% refusal rate on its 48 fixed cases. This is a bounded measurement, not a promise that every possible prompt receives an answer. Learned refusal patterns can still appear, and generated answers can be incorrect. Applications may add their own policies at the client boundary without changing the local artifact.

Voice and vision

Voice and vision are optional signed packages. They remain separate from the main GGUF so text-only users keep the compact primary download.

Vision uses the matching native mmproj-F16 projector and is connected to the OpenAI-compatible image-input route when installed with --with-vision.

Voice contains a compact 0.6B Q8_0 speech-recognition encoder and its matching audio projector. The installer verifies and retains this package independently. The OpenCode audio bridge remains separate from the main image/text endpoint; the card does not claim direct microphone or audio-attachment support until that bridge is released.

Processing long contexts

The release runtime is validated at 8,192 tokens. Larger contexts consume more memory and require separate quality evaluation; they are not part of the v1 release claim.

Best practices

• Use low sampling temperature for code and structured tool arguments.
• Validate tool names and arguments against a declared schema.
• Review generated code and tool actions before execution.
• Give agent clients the minimum required filesystem and network permissions.
• Keep the local endpoint on 127.0.0.1.

Limitations

• The routed candidate is not yet available as one merged standalone GGUF.
• The code route is intentionally isolated because it reduces ordinary English and text quality when used as the default route.
• Routed candidate diagnostics are not official BFCL, SWE-bench, Terminal-Bench, TAU2, LiveCodeBench, GPQA-D, HLE, or MMLU-Pro scores.
• Outputs may contain factual errors, broken code, invalid arguments, or invented citations.
• Search quality depends on the connected search tool and its source coverage.
• Other languages are not covered by the v1 release claim.
• Local execution does not make a tool-enabled client harmless.
• The model is not a substitute for professional medical, legal, financial, or safety-critical review.

Artifact integrity

Artifact	SHA-256
metaflora-incubus-v1.gguf	df850ca6f8d47b2d92db99fb623a36e9d35f3ad7737a588e7ff562ac5229b3fc
incubusctl-darwin-arm64.tar.gz	3582a04a8097787ad73ecfa64ae458083fd5c6188e821f90eef88005d217f79a
Voice package	32c051da39784ef5dea000df4d263f5e4b8d50631207a896c1961e847a22b818
Vision package	5f9ba0fb6af356b60e4082c2747b720f705890d05cdd6dfce8fd89a890b9d935

The repository includes an installer manifest, detached signature, signed evaluation receipts, raw outputs, runtime information, and the legal notices required for distribution.

Reproducibility

• Model revision: 573eb9e077a833f676740c7ba2dca0b34003297d
• Evaluation case-bank SHA-256: 9f18aba6bed35a1165cb5015ab10302f6a219c10ea1e564838a77cc3bcd75d49
• Evaluation runtime SHA-256: ed49acd23b9f537391a97fd3270708cb4cd3240165a57be3d77bfc6bd4fd806
• Seed: 4242
• Temperature: 0

Required attribution and license terms are preserved in THIRD_PARTY_NOTICES.