Rabbit Industries public reference • updated 2026-10-01

Agastya Competition Reference

A detailed specification and benchmark reference for Agastya AGI v01 alongside selected frontier systems. The position column is Rabbit Industries' internal research/reference ordering; it is not a third-party global leaderboard. Agastya's external same-suite performance rank remains pending frozen evaluation and independent replay.

Live Agastya snapshot: verified from current Native evidence on 2026-10-01. Canonical public model remains 1,264,715,776 parameters, 4,096-token context, and 1,000,255,071-token historical/base pretraining snapshot. The F3 successor is a 3,573,746,688-parameter noncanonical candidate with a 4K->32K context curriculum and 8B-token target. Its current runtime state is loaded live below from the authoritative F3 status mirror.

Rank-format correction: every model now uses the same #1–#7 reference-position format. Agastya is displayed as #7, with a separate PUBLIC CANDIDATE / external rank pending status.

Reference position	Model / provider	Access / architecture	Total params	Active params	Context	Training tokens	Public benchmark evidence	Qualification status	Primary source
#1	GPT-6 AstraOpenAI • Sep 3, 2026REFERENCE LEADER	AccessClosed frontierReasoning • coding • computer use • research	TotalUndisclosed	ActiveUndisclosed	Input1,050,000128K max output	PretrainingUndisclosed	ARC-AGI-3 99.9%GPQA Diamond 96.0%Terminal-Bench 4.0 57.9%DeepSWE v1.1 74.1%HLE w/tools 57.2%	PUBLIC FRONTIERSame-suite Agastya parity not yet tested.	OpenAI official
#2	Claude Opus 5.5Anthropic • Sep 22, 2026FRONTIER AGENTIC	AccessClosed frontierAgentic coding • knowledge work • long-running tasks	TotalUndisclosed	ActiveUndisclosed	Standard paid/API1,000,000128K max output \| official Claude Platform default	PretrainingUndisclosed	Terminal-Bench 4.0 66.4%FrontierCode 1.1 Main 54.4%HLE w/tools 67.7%GDPval-AA v2.1 1846 Elo	PUBLIC FRONTIERSame-suite Agastya parity not yet tested.	Anthropic official
#3	Gemini 3.8 FlashGoogle • Sep 2026 stableFRONTIER MULTIMODAL	AccessClosed frontierText • image • video • audio • PDF	TotalUndisclosed	ActiveUndisclosed	Input1,048,57665,536 max output	PretrainingUndisclosed	GPQA Diamond 95.3%DeepSWE v1.1 73.8%Terminal-Bench 4.0 19.1%	PUBLIC FRONTIER	Google official
#4	Qwen3.7-MaxQwen / Alibaba • May 20, 2026FRONTIER AGENT	AccessProprietaryLong-horizon agent foundation	TotalUndisclosed	ActiveUndisclosed	Input1,000,000	PretrainingUndisclosed	GPQA Diamond 92.4%HLE 41.4%HLE w/tools 53.5%Terminal-Bench 2.0 69.7%TB 2.0 is not TB 4.0.	PUBLIC FRONTIER	Qwen official
#5	DeepSeek-V4.1-FlashDeepSeek • Sep 10, 2026FRONTIER AVAILABLE	Architecture552B MoECausal Encoder–Decoder • native multimodal	Total552B	Active8B input / 16B output	Context1,000,000	PretrainingUndisclosed	GPQA Diamond 90.9%HLE 36.8%HLE w/tools 63.9%Terminal-Bench 4.0 31.2%DeepSWE v1.1 74.2%	PUBLIC FRONTIER	DeepSeek official
#6	Llama 4 MaverickMeta • Apr 5, 2025OPEN-WEIGHT REFERENCE	ArchitectureMoE multimodal128 routed experts + shared expert	Total400B	Active17B	Context1M classScout is Meta's 10M-context variant	Family mixture>30TLlama 4 family mixture	Official Meta release publishes broad multimodal, reasoning and coding comparisons.Not placed into the 2026 same-suite matrix where methods differ.	OPEN-WEIGHT	Meta official
#7	Agastya AGI v01Rabbit Industries Pvt Ltd • Bramha Medha Native INDIAPUBLIC CANDIDATE	Canonical v01 architectureDense nativeF3 successor is a separate noncanonical training candidate	Canonical v01 total1.265BExact: 1,264,715,776F3 successor total: 3.574B	Canonical v01 active1.265BExact: 1,264,715,776 • 100% activeF3 successor active: 3.574B • dense	Canonical v01 context4,096Canonical deployed contextF3 live context: 4K • curriculum target: 32K	F3 LIVE TRAININGconnecting…Loading authoritative F3 runtime…Base pretraining: 1,000,255,071	GPQA Diamond PENDINGHLE PENDINGHLE w/tools PENDINGTerminal-Bench 4.0 PENDINGDeepSWE v1.1 PENDINGF3 successor: 3.574B • 4K→32K • target 8B • candidate-onlyLIVE STATUS: connecting…Pipeline: loading…Next: loading…Benchmarks publish automatically after trained-candidate same-suite run + independent replay.	PUBLIC CANDIDATESame-suite qualification pending.	Rabbit Industries official profilePublic specifications + qualification statusNative Brain Evidence R1

Same-benchmark evidence matrix

Benchmark versions are kept explicit. A dash means this page is not asserting a directly verified value from the cited official source. Agastya stays pending rather than receiving an estimated score.

Benchmark	GPT-6 Astra	Claude Opus 5.5	Gemini 3.8 Flash	Qwen3.7-Max	DeepSeek V4.1 Flash	Llama 4 Maverick	Agastya AGI v01
GPQA Diamond	96.0%	—	95.3%	92.4%	90.9%	—	PENDING SAME-SUITE RUN
Terminal-Bench 4.0	57.9%	66.4%	19.1%	— (69.7% is TB 2.0)	31.2%	—	PENDING SAME-SUITE RUN
DeepSWE v1.1	74.1%	—	73.8%	—	74.2%	—	PENDING SAME-SUITE RUN
HLE w/tools	57.2%	67.7%	—	53.5%	63.9%	—	PENDING SAME-SUITE RUN

Agastya disclosure — aligned with the same comparison format

Reference position#7

Same # format as the other entries; separate external-rank status avoids ambiguity.

Canonical v01 parameters1.265B total / active

Exact: 1,264,715,776 total and active parameters. These are fixed architecture properties, not progress counters.

Canonical v01 context4,096 tokens

Canonical deployed context. F3 uses a separate 4K→8K→16K→32K curriculum.

Base pretraining tokens1.000B tokens

Historical/base evidence snapshot: 1,000,255,071 tokens. This is separate from the live F3 successor training counter.

Native lineageRandom initialization

Canonical lineage does not use third-party model weights as initialization.

Benchmark stateQualification in progress

No same-suite frontier score is published until frozen evaluation and independent replay complete.

External rankNot established

External leaderboard acceptance is a future evidence gate.

Company designationThe First World Native Environment

Rabbit Industries designation; independent global-first verification remains pending.

F3 successor parameters3.574B total / active

Exact: 3,573,746,688 total and active parameters. Dense candidate-only architecture; live runtime status loading…

F3 live context / curriculum4K current → 32K target

Context advances only at curriculum stage transitions: 4K → 8K → 16K → 32K. Parameter count stays fixed while weights train.

Methodology: Reference positions are Rabbit Industries' research/reference ordering, not a universal intelligence score. Vendor results can differ by harness, effort level, tool access, safeguards and benchmark version. A direct Agastya performance rank requires identical frozen suites and independent replay.