Post
41
BananaMind Daily Digest is back!
š SupraLabsāÆ/āÆSupra2.6āPico ā a brandānew 2.97āÆMāparameter tiny language model
SupraLabs/Supra2.6-Pico
Why it matters: the model packs 2,972,496 unique parameters into an 11.9āÆMB fp32 weight file, runs with a 1,024ātoken context and a 4,096ātoken vocabulary. Its hybrid architecture mixes three gatedādeltaānet (GDNā2) blocks with a gated groupedāquery attention block per group, giving linearātime recurrence for 6 of 8 layers and global attention for the remaining 2. The design includes learned softmax attentionāresidual mixing, RMSNorm, and SwiGLU FFNs (hidden 352). This is a notable step for subā3āÆM models, offering strong efficiency on edge hardware while introducing novel GDNā2 and gatedāGQA mechanisms.
š¤ BananaMindāÆ/āÆBananaMindā3ā10MāTraining ā updated training repo for the 3ā10āÆMāparameter series
BananaMind/BananaMind-3-10M-Training
Mid-training checkpoints of BananaMind 3 10M
š§Ŗ BananaMindāÆ/āÆ2.5āproātest ā community discussion on the BananaMind 2.5 Pro checkpoint
BananaMind/2.5-pro-test
Early feedback thread evaluating performance and potential issues of the 2.5āpro checkpoint before a formal release.
š OpenāÆASRāÆLeaderboard ā Egyptian and Nigerian English added
https://huggingface.co/blog/open-asr-leaderboard-liva
Expands the multilingual ASR benchmark with two new dialects, improving coverage for African English varieties.
š AxiomicLabs ā OpenāÆSLMāÆLeaderboard discussion space
AxiomicLabs/Open_SLM_Leaderboard
Ongoing conversation about ranking smallālanguage models, inviting submissions and metric proposals.
šŖ NVIDIA CosmosāÆ3 ā Open models for interactive world simulation (article)ā¦
ā
š Posted 06 Nov 2026, 18:55 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.
š SupraLabsāÆ/āÆSupra2.6āPico ā a brandānew 2.97āÆMāparameter tiny language model
SupraLabs/Supra2.6-Pico
Why it matters: the model packs 2,972,496 unique parameters into an 11.9āÆMB fp32 weight file, runs with a 1,024ātoken context and a 4,096ātoken vocabulary. Its hybrid architecture mixes three gatedādeltaānet (GDNā2) blocks with a gated groupedāquery attention block per group, giving linearātime recurrence for 6 of 8 layers and global attention for the remaining 2. The design includes learned softmax attentionāresidual mixing, RMSNorm, and SwiGLU FFNs (hidden 352). This is a notable step for subā3āÆM models, offering strong efficiency on edge hardware while introducing novel GDNā2 and gatedāGQA mechanisms.
š¤ BananaMindāÆ/āÆBananaMindā3ā10MāTraining ā updated training repo for the 3ā10āÆMāparameter series
BananaMind/BananaMind-3-10M-Training
Mid-training checkpoints of BananaMind 3 10M
š§Ŗ BananaMindāÆ/āÆ2.5āproātest ā community discussion on the BananaMind 2.5 Pro checkpoint
BananaMind/2.5-pro-test
Early feedback thread evaluating performance and potential issues of the 2.5āpro checkpoint before a formal release.
š OpenāÆASRāÆLeaderboard ā Egyptian and Nigerian English added
https://huggingface.co/blog/open-asr-leaderboard-liva
Expands the multilingual ASR benchmark with two new dialects, improving coverage for African English varieties.
š AxiomicLabs ā OpenāÆSLMāÆLeaderboard discussion space
AxiomicLabs/Open_SLM_Leaderboard
Ongoing conversation about ranking smallālanguage models, inviting submissions and metric proposals.
šŖ NVIDIA CosmosāÆ3 ā Open models for interactive world simulation (article)ā¦
ā
š Posted 06 Nov 2026, 18:55 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.