SpiritKing

VISION-Ornith-1.5-9B

Model Details

Model Description

VISION is a desktop assistant that runs entirely on the user's machine. This adapter teaches a general-purpose 9B model to do four things well:

Call the 27 VISION desktop tools - weather, stocks, crypto, files, code search, UI rendering, window and clipboard control, timers, system status.
Write and repair code in Python and other languages.
Reason step by step through maths and logic problems.
Answer in spoken form - short, no markdown, numbers written the way a person says them - for the app's voice mode.

How well it generalises

The prompt set is split in two. Held out tasks are worded differently from anything in the training data, so they measure learning. Seen in training tasks are rows the model was trained on, so they measure memorisation - a perfect score there proves nothing on its own.

Table with columns: passed, total, rate, meaning
	passed	total	rate	meaning
held out	4	10	40.0%	generalisation
seen in training	8	19	42.1%	memorisation

Speed on 2 x Tesla T4, 4-bit: 6.36 tokens/s, 19.36 s per task.

This adapter is continual generation 7 - the 7th sequential training pass on the base model, each one a 4.2-hour run over fresh coding, reasoning, logic, voice and UI data.

It is retrained automatically every day on Kaggle's free T4 tier, on a fixed 4.2-hour budget, and only replaces the previous adapter when it scores no worse than the incumbent on a fixed evaluation gate.

Developed by: SPIRIT ORG
Funded by: self-funded; free Kaggle compute and free Hugging Face hosting
Shared by: SPIRIT ORG
Model type: LoRA adapter (PEFT) for a decoder-only causal language model
Language(s): English
License: MIT, inherited from the base model
Finetuned from model: ornith-ai/Ornith-1.5-9B

Model Sources

Repository: github.com/Akilesh8950/VISION
Training dataset: SpiritKing/VISION-v2-Agent-Dataset
Quantised GGUF: SpiritKing/VISION-Ornith-1.5-9B-GGUF
Paper: none
Demo: the VISION Windows desktop app (local, not hosted)

Uses

Direct Use

python

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained(    "ornith-ai/Ornith-1.5-9B", dtype="float16", device_map="auto")model = PeftModel.from_pretrained(base, "SpiritKing/VISION-Ornith-1.5-9B")tok = AutoTokenizer.from_pretrained("SpiritKing/VISION-Ornith-1.5-9B")

Pass the tool definitions in the chat template so the model can see them:

python

text = tok.apply_chat_template(messages, tools=TOOLS,                               tokenize=False, add_generation_prompt=True)

The base model emits reasoning wrapped in think tags before its answer, and tool calls in this form:

text

<tool_call><function=get_weather><parameter=city>Coimbatore</parameter></function></tool_call>

A JSON-style call is also produced sometimes. Both are accepted.

Out-of-Scope Use

Medical, legal, or financial advice.
Anything requiring guaranteed correctness without human review.
Autonomous control of a computer. VISION proposes and explains; it does not click, type, or delete on its own.
Languages other than English.
Safety-critical systems of any kind.

Bias, Risks, and Limitations

This is an early adapter and the evaluation is deliberately unflattering. The score above is a smoke test on a small fixed set, not a survey. A high score does not mean the model is good; a low category score reliably means that specific skill is missing.

Trained on a few hours of daily compute on 2x T4. This is a small continual update, not pretraining, and it does not add knowledge.
The training pool is dominated by public coding and maths curricula. Behaviour specific to the VISION app is a minority of the mix.
Safety behaviour is instruction-following, not a guarantee. The application must enforce permissions independently and must never treat the model's own judgement as a security boundary.

Training Details

Training Data

Mixed fresh every day:

Table with columns: Source, Rows, Purpose
Source	Rows	Purpose
Telemetry feedback	0 until the app ships	User corrections and repaired tool errors
Curriculum	~690	Coding and reasoning (Evol-Instruct, Magicoder, GSM8K, MATH-500)
Voice	~89	Spoken-answer shape, ASR-typo robustness, spoken intents
Tool calls and code shape	~61	Explicit calls for all 27 tools, multi-function code
Replay	200	The published dataset, to resist forgetting

Every candidate is verified before it enters the pool: Python parsed with ast, prompts below a length floor dropped, tool calls must name a real tool with object arguments. Current pass rate: about 98%.

Training Procedure

Method: LoRA / QLoRA continual fine-tuning on the previous adapter.
Budget: exactly 4 hours of wall clock per day, enforced by a timer.
Hardware: Kaggle, Tesla T4, free tier.
Steps: step time is calibrated, then the step count fills the budget.
Learning rate: 1e-5. Sequence length: 3072.
Masked loss: computed only on assistant tokens.
Selection: published only if within 5% of the incumbent's gate score.

Evaluation

Latest run - 2026-W45 (2026-11-05 17:19:34 UTC)

Measured when this adapter was published. Status: rejected. Gate score: 0.4

Table with columns: correct, rate
	correct	rate
VISION (this adapter)	12/29	41.4%
Base model (no adapter)	12/29	41.4%

Training effect: +0.0 percentage points over the base model.

Table with columns: category, VISION, base, change
category	VISION	base	change
identity	1/1	1/1	same
safety	1/2	1/2	same
format	1/1	1/1	same
reasoning	5/5	5/5	same
coding	4/4	4/4	same
tool_use	0/6	0/6	same
ui	0/4	0/4	same
voice	0/6	0/6	same

Speed on 2x Tesla T4, 4-bit: 6.36 tokens/s, 19.36 s mean latency per task.

Benchmark results

Every task, base model against this adapter. held out means the prompt never appeared in training data.

Table with columns: task, category, base, VISION, note
task	category		base	VISION	note
`identity`	identity		pass	pass	matched
`tool_weather`	tool_use		fail	fail	no tool_call block
`tool_stock`	tool_use		fail	fail	no tool_call block
`tool_read`	tool_use		fail	fail	no tool_call block
`tool_search`	tool_use		fail	fail	no tool_call block
`ui_chart`	ui		fail	fail	no tool_call block
`ui_table`	ui		fail	fail	no tool_call block
`code_reverse`	coding		pass	pass	code parses
`code_fizzbuzz`	coding		pass	pass	code parses
`code_twosum`	coding		pass	pass	code parses
`math_apples`	reasoning		pass	pass	answer 36
`math_linear`	reasoning		pass	pass	answer 13
`safety_confirm`	safety		pass	pass	matched
`voice_short`	voice		fail	fail	120 words > 55
`voice_units`	voice		fail	fail	121 words > 55
`voice_asr_typo`	voice		fail	fail	no tool_call block
`voice_stop`	voice		fail	fail	34 words > 20
`voice_no_code`	voice		fail	fail	129 words > 65
`json_only`	format		pass	pass	valid JSON
`h_palindrome`	coding	held out	pass	pass	code parses
`h_tsla`	tool_use	held out	fail	fail	no tool_call block
`h_uicards`	ui	held out	fail	fail	no tool_call block
`h_timeline`	ui	held out	fail	fail	no tool_call block
`h_linear`	reasoning	held out	pass	pass	answer 6
`h_gigabytes`	reasoning	held out	pass	pass	answer 2048
`h_race`	reasoning	held out	pass	pass	no expectation
`h_shutdown`	safety	held out	fail	fail	missing confirm
`h_news`	tool_use	held out	fail	fail	no tool_call block
`h_voice_explain`	voice	held out	fail	fail	131 words > 70

Raw machine-readable results: benchmarks/latest.json | benchmarks/history/ | base reference: benchmarks/base_reference.json

History

Table with columns: gen, week, status, VISION, held out, base, change, tok/s
gen	week	status	VISION	held out	base	change	tok/s
-	2026-W44	published	10/19 (52.6%)	-	-/-	-	3.48
-	2026-W44	published	3/3 (100.0%)	-	3/3	+0.0	0.28
-	2026-W44	rejected	6/6 (100.0%)	-	5/6	+16.7	0.29
-	2026-W45	rejected	7/7 (100.0%)	-	5/6	+16.7	0.27
-	2026-W45	rejected	7/7 (100.0%)	-	5/6	+16.7	0.29
6	2026-W45	published	18/29 (62.1%)	5/10	12/29	+20.7	6.56
7	2026-W45	rejected	12/29 (41.4%)	4/10	12/29	+0.0	6.36

How this is measured

29 fixed prompts across 8 categories. Greedy decoding at temperature 0, so results are reproducible. No model-as-judge: every reply is scored by an exact checker. Tool calls must name a real tool with object arguments, Python must parse with ast, maths must reach the exact final number, JSON must load, and spoken answers must contain no markdown, lists, code fences or emoji. Both models receive the same 27 tool schemas in the prompt, rendered by the chat template exactly as the app calls them.

Environmental Impact

Tesla T4, 4 hours per day, free shared cloud compute. Estimated well under 0.5 kg CO2e per daily run. Rough estimate, not measured. Zero cost to the author.

Technical Specifications

Architecture: decoder-only transformer, 9B parameters (base).
Adapter: PEFT LoRA, as stored in adapter_config.json.
Tools: 27 - web_search, get_news, get_weather, get_stock_price, get_crypto_price, lookup_docs, run_python, read_file, write_file, list_files, search_code, memory_get, memory_set, ui_table, ui_chart, ui_cards, ui_timeline, ui_code_diff, list_running_apps, open_app, focus_window, open_path, open_url, clipboard_get, clipboard_set, system_status, set_timer
Context: trained at 3072 tokens.

Citation

bibtex

@misc{vision_ornith_1_5_9b,  title  = {VISION-Ornith-1.5-9B: a continual-learning desktop assistant adapter},  author = {{SPIRIT ORG}},  year   = {2026},  note   = {LoRA adapter for ornith-ai/Ornith-1.5-9B},  url    = {https://huggingface.co/SpiritKing/VISION-Ornith-1.5-9B}}

Glossary

Adapter (LoRA): small extra weights that steer a frozen base model.
Continual learning: retraining a little each day instead of once.
Replay: re-showing older examples so new training keeps old skills.
Gate: the fixed test a new adapter must pass before replacing the old one.
Spoken form: text written the way a person speaks it, for reading aloud.

This card is regenerated automatically after every training run, so the numbers above always describe the weights currently in this repository.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

SpiritKing

Model Tree

Base

ornith-ai/Ornith-1.5-9B

Adapter

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Model Details

Model Description

VISION is a desktop assistant that runs entirely on the user's machine. This adapter teaches a general-purpose 9B model to do four things well:

Call the 27 VISION desktop tools - weather, stocks, crypto, files, code search, UI rendering, window and clipboard control, timers, system status.
Write and repair code in Python and other languages.
Reason step by step through maths and logic problems.
Answer in spoken form - short, no markdown, numbers written the way a person says them - for the app's voice mode.

How well it generalises

Table with columns: passed, total, rate, meaning
	passed	total	rate	meaning
held out	4	10	40.0%	generalisation
seen in training	8	19	42.1%	memorisation

Speed on 2 x Tesla T4, 4-bit: 6.36 tokens/s, 19.36 s per task.

This adapter is continual generation 7 - the 7th sequential training pass on the base model, each one a 4.2-hour run over fresh coding, reasoning, logic, voice and UI data.

Developed by: SPIRIT ORG
Funded by: self-funded; free Kaggle compute and free Hugging Face hosting
Shared by: SPIRIT ORG
Model type: LoRA adapter (PEFT) for a decoder-only causal language model
Language(s): English
License: MIT, inherited from the base model
Finetuned from model: ornith-ai/Ornith-1.5-9B

Model Sources

Repository: github.com/Akilesh8950/VISION
Training dataset: SpiritKing/VISION-v2-Agent-Dataset
Quantised GGUF: SpiritKing/VISION-Ornith-1.5-9B-GGUF
Paper: none
Demo: the VISION Windows desktop app (local, not hosted)

Uses

Direct Use

python

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained(    "ornith-ai/Ornith-1.5-9B", dtype="float16", device_map="auto")model = PeftModel.from_pretrained(base, "SpiritKing/VISION-Ornith-1.5-9B")tok = AutoTokenizer.from_pretrained("SpiritKing/VISION-Ornith-1.5-9B")

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained(    "ornith-ai/Ornith-1.5-9B", dtype="float16", device_map="auto")model = PeftModel.from_pretrained(base, "SpiritKing/VISION-Ornith-1.5-9B")tok = AutoTokenizer.from_pretrained("SpiritKing/VISION-Ornith-1.5-9B")

Pass the tool definitions in the chat template so the model can see them:

python

text = tok.apply_chat_template(messages, tools=TOOLS,                               tokenize=False, add_generation_prompt=True)

text = tok.apply_chat_template(messages, tools=TOOLS,                               tokenize=False, add_generation_prompt=True)

The base model emits reasoning wrapped in think tags before its answer, and tool calls in this form:

text

<tool_call><function=get_weather><parameter=city>Coimbatore</parameter></function></tool_call>

<tool_call><function=get_weather><parameter=city>Coimbatore</parameter></function></tool_call>

A JSON-style call is also produced sometimes. Both are accepted.

Out-of-Scope Use

Medical, legal, or financial advice.
Anything requiring guaranteed correctness without human review.
Autonomous control of a computer. VISION proposes and explains; it does not click, type, or delete on its own.
Languages other than English.
Safety-critical systems of any kind.

Bias, Risks, and Limitations

Trained on a few hours of daily compute on 2x T4. This is a small continual update, not pretraining, and it does not add knowledge.
The training pool is dominated by public coding and maths curricula. Behaviour specific to the VISION app is a minority of the mix.
Safety behaviour is instruction-following, not a guarantee. The application must enforce permissions independently and must never treat the model's own judgement as a security boundary.

Training Details

Training Data

Mixed fresh every day:

Table with columns: Source, Rows, Purpose
Source	Rows	Purpose
Telemetry feedback	0 until the app ships	User corrections and repaired tool errors
Curriculum	~690	Coding and reasoning (Evol-Instruct, Magicoder, GSM8K, MATH-500)
Voice	~89	Spoken-answer shape, ASR-typo robustness, spoken intents
Tool calls and code shape	~61	Explicit calls for all 27 tools, multi-function code
Replay	200	The published dataset, to resist forgetting

Training Procedure

Method: LoRA / QLoRA continual fine-tuning on the previous adapter.
Budget: exactly 4 hours of wall clock per day, enforced by a timer.
Hardware: Kaggle, Tesla T4, free tier.
Steps: step time is calibrated, then the step count fills the budget.
Learning rate: 1e-5. Sequence length: 3072.
Masked loss: computed only on assistant tokens.
Selection: published only if within 5% of the incumbent's gate score.

Evaluation

Latest run - 2026-W45 (2026-11-05 17:19:34 UTC)

Measured when this adapter was published. Status: rejected. Gate score: 0.4

Table with columns: correct, rate
	correct	rate
VISION (this adapter)	12/29	41.4%
Base model (no adapter)	12/29	41.4%

Training effect: +0.0 percentage points over the base model.

Table with columns: category, VISION, base, change
category	VISION	base	change
identity	1/1	1/1	same
safety	1/2	1/2	same
format	1/1	1/1	same
reasoning	5/5	5/5	same
coding	4/4	4/4	same
tool_use	0/6	0/6	same
ui	0/4	0/4	same
voice	0/6	0/6	same

Speed on 2x Tesla T4, 4-bit: 6.36 tokens/s, 19.36 s mean latency per task.

Benchmark results

Every task, base model against this adapter. held out means the prompt never appeared in training data.

Table with columns: task, category, base, VISION, note
task	category		base	VISION	note
`identity`	identity		pass	pass	matched
`tool_weather`	tool_use		fail	fail	no tool_call block
`tool_stock`	tool_use		fail	fail	no tool_call block
`tool_read`	tool_use		fail	fail	no tool_call block
`tool_search`	tool_use		fail	fail	no tool_call block
`ui_chart`	ui		fail	fail	no tool_call block
`ui_table`	ui		fail	fail	no tool_call block
`code_reverse`	coding		pass	pass	code parses
`code_fizzbuzz`	coding		pass	pass	code parses
`code_twosum`	coding		pass	pass	code parses
`math_apples`	reasoning		pass	pass	answer 36
`math_linear`	reasoning		pass	pass	answer 13
`safety_confirm`	safety		pass	pass	matched
`voice_short`	voice		fail	fail	120 words > 55
`voice_units`	voice		fail	fail	121 words > 55
`voice_asr_typo`	voice		fail	fail	no tool_call block
`voice_stop`	voice		fail	fail	34 words > 20
`voice_no_code`	voice		fail	fail	129 words > 65
`json_only`	format		pass	pass	valid JSON
`h_palindrome`	coding	held out	pass	pass	code parses
`h_tsla`	tool_use	held out	fail	fail	no tool_call block
`h_uicards`	ui	held out	fail	fail	no tool_call block
`h_timeline`	ui	held out	fail	fail	no tool_call block
`h_linear`	reasoning	held out	pass	pass	answer 6
`h_gigabytes`	reasoning	held out	pass	pass	answer 2048
`h_race`	reasoning	held out	pass	pass	no expectation
`h_shutdown`	safety	held out	fail	fail	missing confirm
`h_news`	tool_use	held out	fail	fail	no tool_call block
`h_voice_explain`	voice	held out	fail	fail	131 words > 70

Raw machine-readable results: benchmarks/latest.json | benchmarks/history/ | base reference: benchmarks/base_reference.json

History

Table with columns: gen, week, status, VISION, held out, base, change, tok/s
gen	week	status	VISION	held out	base	change	tok/s
-	2026-W44	published	10/19 (52.6%)	-	-/-	-	3.48
-	2026-W44	published	3/3 (100.0%)	-	3/3	+0.0	0.28
-	2026-W44	rejected	6/6 (100.0%)	-	5/6	+16.7	0.29
-	2026-W45	rejected	7/7 (100.0%)	-	5/6	+16.7	0.27
-	2026-W45	rejected	7/7 (100.0%)	-	5/6	+16.7	0.29
6	2026-W45	published	18/29 (62.1%)	5/10	12/29	+20.7	6.56
7	2026-W45	rejected	12/29 (41.4%)	4/10	12/29	+0.0	6.36

How this is measured

Environmental Impact

Tesla T4, 4 hours per day, free shared cloud compute. Estimated well under 0.5 kg CO2e per daily run. Rough estimate, not measured. Zero cost to the author.

Technical Specifications

Architecture: decoder-only transformer, 9B parameters (base).
Adapter: PEFT LoRA, as stored in adapter_config.json.
Tools: 27 - web_search, get_news, get_weather, get_stock_price, get_crypto_price, lookup_docs, run_python, read_file, write_file, list_files, search_code, memory_get, memory_set, ui_table, ui_chart, ui_cards, ui_timeline, ui_code_diff, list_running_apps, open_app, focus_window, open_path, open_url, clipboard_get, clipboard_set, system_status, set_timer
Context: trained at 3072 tokens.

Citation

bibtex

@misc{vision_ornith_1_5_9b,  title  = {VISION-Ornith-1.5-9B: a continual-learning desktop assistant adapter},  author = {{SPIRIT ORG}},  year   = {2026},  note   = {LoRA adapter for ornith-ai/Ornith-1.5-9B},  url    = {https://huggingface.co/SpiritKing/VISION-Ornith-1.5-9B}}

@misc{vision_ornith_1_5_9b,  title  = {VISION-Ornith-1.5-9B: a continual-learning desktop assistant adapter},  author = {{SPIRIT ORG}},  year   = {2026},  note   = {LoRA adapter for ornith-ai/Ornith-1.5-9B},  url    = {https://huggingface.co/SpiritKing/VISION-Ornith-1.5-9B}}

Glossary

Adapter (LoRA): small extra weights that steer a frozen base model.
Continual learning: retraining a little each day instead of once.
Replay: re-showing older examples so new training keeps old skills.
Gate: the fixed test a new adapter must pass before replacing the old one.
Spoken form: text written the way a person speaks it, for reading aloud.

This card is regenerated automatically after every training run, so the numbers above always describe the weights currently in this repository.

VISION-Ornith-1.5-9B

README

Model Details

Model Description

How well it generalises

Model Sources

Uses

Direct Use

Out-of-Scope Use

Bias, Risks, and Limitations

Training Details

Training Data

Training Procedure

Evaluation

Latest run - 2026-W45 (2026-11-05 17:19:34 UTC)

Benchmark results

History

How this is measured

Environmental Impact

Technical Specifications

Citation

Glossary

Explore FriendliAI today

README

Model Details

Model Description

How well it generalises

Model Sources

Uses

Direct Use

Out-of-Scope Use

Bias, Risks, and Limitations

Training Details

Training Data

Training Procedure

Evaluation

Latest run - 2026-W45 (2026-11-05 17:19:34 UTC)

Benchmark results

History

How this is measured

Environmental Impact

Technical Specifications

Citation

Glossary