Records found: 6
Cerebras released MiniMax-M2-REAP-162B-A10B, a REAP-pruned SMoE checkpoint that trims experts to reduce memory while keeping near-original performance on code, reasoning and agentic tool calls.
'Ling 2.0 is a reasoning-first sparse MoE family from Ant Group that keeps per-token compute low with a 1/32 activation recipe while scaling from 16B to 1T parameters.'
Glyph converts ultra-long text into page images processed by a VLM to achieve 3–4× effective token compression and roughly 4× faster prefill and decoding on 128K inputs.
'GLM-4.6 expands context to 200K tokens, reduces token consumption on real-world coding benchmarks, and offers open weights for local inference and research.'
'A concise 2025 guide to AI agents covering what they are, where they work reliably, risks, architecture patterns, and evaluation strategies.'
TII’s Falcon-H1 series introduces hybrid Transformer-SSM models that combine efficiency and performance, supporting long contexts, multilingual processing, and scalable deployment from 0.5B to 34B parameters.