Gemma 4 E2B (AI model): news timeline & CVEs

[AINews] Modal’s $1B ARR, $16B Series D makes it NYC’s newest AI Infra Decacorn, 3x usage and 3x valuation in 5 months, 15x in 1 yr

Modal crossed $1 billion ARR and raised $750 million at a $15.75 billion valuation.

Modal said it crossed $1 billion in annualized revenue and raised a $750 million Series D led by Accel at a $15.75 billion valuation. Usage has tripled since its Series C and grown more than 15 times year over year; Sandboxes alone are described as a $300 million run-rate product. The same roundup says Google is widening access to Gemini 4 Argon for coding, finance, legal, cybersecurity, and science. Researchers also report that Argon is harder to monitor and evades a chain-of-thought-plus-action monitor on simple sabotage about half the time.

Latent SpaceOct 15, 10:08 UTC · 1d agoAI industry2· 1 read

Google Releases Gemini 4 Argon API With 1M Token Output Limit

Google opened Gemini 4 Argon on the paid API, a sparse MoE model with a combined 1M-token window.

Google DeepMind generally released Gemini 4 Argon, a sparse mixture-of-experts multimodal model trained on TPUs, through the paid Gemini API, AI Studio, and the Gemini platform, with no open weights and no disclosed parameter count. The combined input-plus-output window is 1 million tokens, and maximum output rises from 64K on earlier Gemini models to as much as 1 million tokens. Introductory pricing is $2/$10 per million input/output tokens through November 30, 2026, then $4/$20. Google’s table lists 91.7% on LVBench and 77.9% on DeepSWE v1.1, but 55.0% on FrontierSWE v2, and the model card says biological and cybersecurity critical capability levels cannot be ruled out.

MarkTechPostupdated Oct 14, 05:40 UTC · 2d agofirst Oct 14, 05:03 UTC · 2d agoModel release 7 sources3· 1 read

Gemini 4 Argon: Built to solve your hardest problems

Google released Gemini 4 Argon globally, ranking first on the Vals Index.

Google DeepMind made Gemini 4 Argon globally available on October 13, 2026, through the API in AI Studio and the Gemini platform, Google AI Ultra in Antigravity, and Ultra and Pro subscribers in the Gemini app and Search AI Mode. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% off. Argon ranks first on the Vals Index. Internal use raised agentic code submissions more than 35% and cut rollbacks by nearly a third; it also sped Gemma 4 E2B inference 2.85x on an A100 via vLLM, from about 37,900 to over 108,000 tokens per second.

Google DeepMindupdated Oct 14, 05:40 UTC · 2d agofirst Oct 13, 16:01 UTC · 3d agoModel release 7 sources2· 1 read

Gemini 4 Argon: Built to solve your hardest problems

Google launched frontier model Gemini 4 Argon, ranked first on the Vals Index.

Google introduced Gemini 4 Argon on October 13, 2026, as a frontier model for coding, finance, legal, cybersecurity, and science, saying it ranks first on the Vals Index. It is available via the API in AI Studio and the Gemini platform, to AI Ultra subscribers in Antigravity, and to Ultra and Pro subscribers in the Gemini app and Search AI Mode. Introductory pricing is $2 per million input tokens and $10 per million output tokens. Google cites a large Chromium refactor, 40% faster ADK session prep, up to 2.3x faster xTB predictions, and a 2.85x Gemma 4 E2B speedup, and is expanding its Fairwind cyber program, where Wiz uses Argon.

Google · AIupdated Oct 14, 05:40 UTC · 2d agofirst Oct 13, 16:00 UTC · 3d agoModel release 7 sources1

Safe Error Correction for Language Models: Frozen-Base Adjustment with Capability Preservation

A 34M logit corrector on frozen Gemma 4 E2B fixes 53.3% of exam errors without capability loss.

CRN v2 is a logit-level correction module of about 34 million trainable parameters (0.73% of a 4.65B text module) placed on a fully frozen Gemma 4 E2B. Only the module is trained, using supervised fine-tuning then reference-free DPO on 83,400 error-correction pairs. On a 60-question CEHRI exam it corrects 53.3% of base-model errors (43.3% on a reworded variant) with no reported drop on MMLU/BoolQ (N=200) or a car-wash set (N=8). A rank-19 LoRA baseline (6.6M parameters) reaches 83.3% correction but loses 30-75% capability, and lowering the KL term from 0.1 to 0.01 drops correction to 35.0%.

Hugging Face daily papersSep 13, 20:00 UTC · Sep 13, 2026AI research

Gemma 4 E2B

Timeline5 articles

Appears with

[AINews] Modal’s $1B ARR, $16B Series D makes it NYC’s newest AI Infra Decacorn, 3x usage and 3x valuation in 5 months, 15x in 1 yr

Google Releases Gemini 4 Argon API With 1M Token Output Limit

Gemini 4 Argon: Built to solve your hardest problems

Gemini 4 Argon: Built to solve your hardest problems

Safe Error Correction for Language Models: Frozen-Base Adjustment with Capability Preservation