GigaChat 3.5 Reasoning lands under MIT — and beats DeepSeek on math
Sber's GigaChat team has open-weighted a 432-billion-parameter reasoning model under an MIT licence, and the numbers put it within touching distance of DeepSeek V4 Flash. GigaChat 3.5 Reasoning runs 28 billion active parameters, trains natively in FP8, and mixes multi-head latent attention with GatedDeltaNet linear-attention layers — a hybrid aimed squarely at long-context cost — with a 262K-token window and three speculative-decoding heads bolted on. The release includes FP8 weights for inference, a bf16 checkpoint for fine-tuning, and GGUF builds, all on Hugging Face.
The training recipe is the interesting part: rather than one RL run, the team trained six domain experts separately — STEM, code, code agent, general agent, dialogue, and "soft skills" — each with its own reward signal, then merged them into one model via on-policy distillation. Tasks the model already solved more than 75% of the time were dropped from the pool as training progressed, so the curriculum gets harder as the model gets better. On the resulting benchmarks it beats DeepSeek V4 Flash Preview on AIME 2025 (89.0 vs 88.95) and AIME 2026 (92.0 vs 90.4), on instruction following (IFBench 77 vs 73.3) and on structured output, while losing the average (68.9 vs 72.7) and losing badly on agentic coding — SWE-bench Verified 64.7 against 78.6, Terminal-Bench 2 at 30.3 against 56.6. It also uses 37% fewer reasoning tokens across the maths evals. Our read: this is a math-and-format win being marketed as a general one; the agentic gap is the number that matters in 2026.
Nvidia and Palantir are wiring Nvidia's own supply chain into an AI stack the two will now sell to everyone else. The collaboration drops Nvidia's open Nemotron models into Palantir Foundry and AIP, grounded in Palantir's Ontology, with Nvidia's cuOpt handling constraint modelling and scenario planning. The first deployment isn't a customer — it's Nvidia, whose Vera Rubin racks each draw on 1.3 million parts across thousands of suppliers, and where the pitch is spotting material constraints earlier and allocating stock by end-to-end production impact. Enterprises post-train Nemotron on their own operational data, and every recommendation and outcome feeds back through Palantir Autopilot and Nvidia's NeMo libraries into a "governed learning loop." The framing is sovereignty: open weights plus on-prem deployment equals control, which is the same argument Mistral's Arthur Mensch has been making about proprietary labs sitting front-row to your business processes. Alex Karp called Nvidia's chain "arguably the most valuable, intricate and complex supply chain in the world." The proof will be whether it survives contact with a company that isn't Nvidia.
The Bank for International Settlements has formally put the AI boom on the financial-stability risk list. BIS head Pablo Hernández de Cos, speaking at a conference hosted by India's central bank, said the world's five largest tech firms will invest more than $1 trillion in AI across 2025 and 2026, while industry forecasts put global AI investment on a path from roughly $500 billion today to as much as $4 trillion by 2030. His specific concern isn't the spending — it's the funding: the buildout is increasingly financed through debt and private credit rather than corporate earnings, and much of that financing is, in his words, opaque and interconnected. Add lofty valuations and market concentration and you get vulnerabilities if profits miss expectations. He drew the parallel himself to railway mania and the dotcom surge, while stressing he isn't predicting a crash. On the upside he cited task-level productivity gains of 10% to 65%, mostly in coding, consulting and professional writing, and estimated AI could lift total factor productivity growth by about half a percentage point a year. Half a point is real money over a decade — and it is a fraction of what $4 trillion in capex needs it to be.
What to watch: whether any regulator actually asks the private-credit lenders financing AI data centres to show their books.
If a 432B model lands under MIT with math scores at the frontier, does the agentic gap still justify closed labs — or is it the last thing they hold? Tell us in the comments.
