{"id":"nr-b13c0319-gigachat-3-5-reasoning-opens-a-large-sparse-model-for-inspection","slug":"gigachat-3-5-reasoning-opens-a-large-sparse-model-for-inspection","title":"GigaChat 3.5 Reasoning Opens a Large Sparse Model for Inspection","observed_at":"2026-09-13T11:14:17.693Z","record_date":"2026-09-13","date_kind":"observed_at","novelty":"new","source_platform":"huggingface.co","topics":["models"],"entities":["Hugging Face"],"related_patterns":[],"source_url":"https://huggingface.co/ai-sage/GigaChat3.5-432B-A28B-Reasoning","source_urls":["https://huggingface.co/ai-sage/GigaChat3.5-432B-A28B-Reasoning"],"basket":{"id":"0a002c08-692d-4485-b9dc-0ffc04cebcca","revision":3,"review_ref":"b13c0319","cluster_id":"6040e2ce-2fbb-4a21-875e-904030f3f27a","mention_count":5},"type":"post","description":"Sber's model-card release exposes weights, architecture choices, and a reasoning-focused training direction through Hugging Face.","why_it_matters":"Evaluation should pin the weights, inference runtime, prompts, hardware, quantization, and tool schema before comparing the model with closed services. The evidence is the official model card and repository; vendor-reported capabilities do not establish independent quality, safety, or cost leadership. Watch for reproducible benchmarks, deployment recipes, failure analysis, and adoption evidence across real Russian-language agent workflows.","verification_level":"source-inspected-local","signal_type":"field_note","schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:gigachat-3-5-reasoning-opens-a-large-sparse-model-for-inspection","retrieval_nugget":"Sber's model-card release exposes weights, architecture choices, and a reasoning-focused training direction through Hugging Face. GigaChat 3.5 Reasoning adds a large open-weight reasoning model to the deployment choices available for Russian and multilingual workloads. This matters now because open weights let teams test language quality, tool behavior, serving cost, and private deployment under their own task contract. The model card","status":"published","visuals":[],"editorial_provenance":{"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:5e0a43cad265b95f59114d915d243d005c4f546f6929fca76adefbbedda89d8f","reviewed_at":"2026-09-13T11:14:17.693Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":2,"implication_count":1,"watch_condition_count":1,"related_record_count":0},"analysis":{"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"GigaChat 3.5 Reasoning adds a large open-weight reasoning model to the deployment choices available for Russian and multilingual workloads.","observed_facts":[{"text":"The model card describes the release as GigaChat's first reasoning model trained with online reinforcement learning.","source_urls":["https://huggingface.co/ai-sage/GigaChat3.5-432B-A28B-Reasoning"]},{"text":"The repository provides FP8 weights and identifies a mixture-of-experts architecture with multilingual, long-context, tool-use, and structured-output capabilities.","source_urls":["https://huggingface.co/ai-sage/GigaChat3.5-432B-A28B-Reasoning"]}],"mechanism":"The mechanism combines sparse expert activation with reasoning training so a large parameter pool can be served without activating every parameter for each token.","why_now":"This matters now because open weights let teams test language quality, tool behavior, serving cost, and private deployment under their own task contract.","implications":["Evaluation should pin the weights, inference runtime, prompts, hardware, quantization, and tool schema before comparing the model with closed services."],"evidence_boundary":"The evidence is the official model card and repository; vendor-reported capabilities do not establish independent quality, safety, or cost leadership.","watch_conditions":["Watch for reproducible benchmarks, deployment recipes, failure analysis, and adoption evidence across real Russian-language agent workflows."],"related_records":[],"new_branch_reason":"No existing New Runtime analytical record covers this exact management and technical mechanism closely enough."},"routes":{"html":"https://newruntime.com/posts/gigachat-3-5-reasoning-opens-a-large-sparse-model-for-inspection/","markdown":"https://newruntime.com/posts/gigachat-3-5-reasoning-opens-a-large-sparse-model-for-inspection.md","json":"https://newruntime.com/posts/gigachat-3-5-reasoning-opens-a-large-sparse-model-for-inspection.json"},"editor":{"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"next_reads":[{"type":"topic","path":"/topics/models/","url":"https://newruntime.com/topics/models/","title":"Models","reason":"Explore the models topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/models.md","json_url":"https://newruntime.com/topics/models.json"},{"type":"collection","path":"/posts/","url":"https://newruntime.com/posts/","title":"Field Notes","reason":"Open the latest editorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/posts.json"},{"type":"collection","path":"/source-ledger.json","url":"https://newruntime.com/source-ledger.json","title":"Source Ledger","reason":"Inspect the public source evidence graph.","media_type":"application/json"}]}