Internal Productllm
Continuum1-9B
Long-context foundation model with hybrid linear attention — open weights on Hugging Face.
Params: ~8.6BContext: 2M tokensMMLU: ~75% (protocol)
LLM research program
Continuous context. Linear compute.
Continuum1-9B is Innomium foundation model for extreme long-context reasoning.
Research-grade evaluationProduction-minded engineeringResponsible handover
Delivery layers, public artifacts, and evaluation notes for this program. Published metrics follow Innomium protocols — validate on your own data before production use.
Delivery layers
From evaluation through inference — how this program is built and shipped.
1
Architecture
Hybrid GLA + Gated DeltaNet with NOPE.
2
Training
Structural distillation + 2M-context expansion.
3
Inference Kernels
continuum-flash-linear-attention Triton kernels.
4
Applications
Reasoning, math, and long-document QA at 2M context.
Why Innomium LLM
2M native context
Linear compute at million-token scale.
Strong benchmarks
75% MMLU, 37% MATH-500.
Open weights
Continuum1-9B on Hugging Face.
Custom kernels
Innomium flash-linear-attention fork.
LLM releases
LLM Arena challenge history
Closed challenge families from our public Arena program. Browse live Arena listings for current rules and leaderboards.
closedllm
Continuum Kernel Optimization Round 1
Completed Jun 2026 · 41 participants · 3 PRs merged
View Arena challengesLLM insights
View all updatesFrequently Asked Questions
~8.6B params, 2M context, BF16 on Hugging Face.
Partner with Innomium on LLM
Whether you need evaluation, adaptation, or production integration — start with the operating constraint.