Head to head: Google: Gemini 3.7 Flash vs DeepSeek-V4-Flash

Google: Gemini 3.7 Flash: **Google Releases Gemini 3.7 Flash Just Three Weeks After Predecessor** *The rapid update introduces lower introductory pricing and claims benchmark gains across coding and workflow automation.* Google is rolling out Gemini 3.7 Flash to replace Gemini 3.6 Flash, replacing the prior model a mere three weeks after its debut. The rapid turnover highlights an aggressive release cadence as Google bypasses the long-anticipated Gemini 3.5 Pro to deploy what it describes as a developer-focused "workhorse" model. The abrupt replacement appears driven by competitive market pressure and developer feedback. Alongside targeted optimizations for coding and agentic tasks, Google is offering Gemini 3.7 Flash at a lower introductory price to counter less expensive rival models. According to details reported by *Ars Technica*, Google Senior Director Tulsee Doshi claimed that Gemini 3.7 Flash delivers noticeable performance gains over version 3.6, particularly in software development and structured workflows. However, these figures represent company-provided benchmark evaluations rather than independently verified third-party testing. In the company-reported metrics: * **FrontierCode 1.1 Main:** Rose to 43.6 percent, up from 34.4 percent on 3.6 Flash. * **DeepSWE v1.1:** Increased to 65.3 percent from 49 percent. * **WebDev Arena:** Climbed to a score of 1,588, compared to 1,538 previously. * **GDP.pdf:** Reached 34 percent, up from 22 percent on the complex document processing benchmark. * **AutomationBench:** Scored 30.4 percent on common business workflow execution, up from 17 percent on 3.6 Flash. While Google's reported metrics indicate significant progress in code generation and document analysis, independent verification will determine how closely these benchmark improvements reflect practical, everyday developer use.

DeepSeek-V4-Flash: **Google Rushes Out Gemini 3.7 Flash Just Weeks After Predecessor, Touting Coding Gains** **Prompt replacement of 3.6 Flash signals accelerated AI iteration as Google aims to undercut rivals on price while boosting agentic performance.** **Aug. 13, 2026** — Google today released Gemini 3.7 Flash, a new iteration of its “workhorse” model that replaces the prior 3.6 Flash version launched only three weeks ago, according to an announcement by Senior Director Tulsee Doshi. The rapid cadence marks an acceleration in Google’s model release cycle, with the company positioning the update as a response to core optimizations and developer feedback rather than a major architectural overhaul. Doshi indicated that the new model offers improved coding and agentic capabilities while arriving at a lower “introductory price,” a move aimed at competing with lower-cost alternatives from rivals. Per company-reported figures, Gemini 3.7 Flash shows notable gains on several benchmarks. On the FrontierCode 1.1 Main test, which measures code generation ability, Google cites an increase to 43.6 percent from 3.6 Flash’s 34.4 percent. The DeepSWE v1.1 evaluation, focused on software engineering tasks, rose to 65.3 percent from 49 percent. On the WebDev Arena, a measure of web development capabilities, the company reports a score of 1,588, up from 1,538. Document comprehension, as tested by the GDP.pdf benchmark, improved to 34 percent versus 3.6 Flash’s 22 percent. AutomationBench, which assesses business workflow execution, saw Google’s reported figure jump to 30.4 percent from 17 percent. These results remain company-supplied claims and have not been independently verified by third-party evaluators at the time of reporting. The accelerated replacement of Gemini 3.6 Flash raises questions about model stability and the pace of Google’s deployment pipeline. The shift could benefit developers seeking incremental improvements without waiting for a major release, but also introduces churn for teams that built workflows around the previous version just weeks ago—particularly given that Google’s documentation notes that temperature, top_p, and top_k parameters are now “deprecated and ignored” in the latest API models, though the specific impact of this change on Gemini 3.7 Flash is not established. Ars Technica first reported the release. Google has not disclosed a timeline for the previously expected Gemini 3.5 Pro model.

Google: Gemini 3.7 Flash: **Headline:** MiniMax Unveils H3 Omni-Modal Model for Multimodal Understanding and 2K Video **Dek:** The organization characterizes H3 as an open model capable of generating 15-second videos with native stereo audio. AI developer MiniMax has introduced MiniMax H3, describing the system as a general-purpose omni-modal generation model engineered to break boundaries between tasks and modalities. According to statements released by MiniMax Research, the H3 model is capable of jointly understanding multimodal contexts spanning text, images, video, and audio. On the generation side, the company claims H3 can produce video with native stereo audio at resolutions of up to 2K and durations of up to 15 seconds. Editorial review of the release materials identifies an unverified date discrepancy across the company's communication channels. MiniMax Research published its statement on July 30, 2026, at 7:00 PM CT, declaring that it was officially launching H3 on that date. However, the primary official MiniMax account shared a link to the announcement titled "MiniMax H3: An open model breaking the boundaries between tasks and modalities" over a day later, on August 1, 2026, at 12:50 AM CT. Before publication can be fully certified, several underlying elements require further independent verification. Although MiniMax bills H3 as an "open model" that has officially launched, the provided materials do not include repository links, model weight access details, open-source software licenses, or public availability confirmations. To confirm the launch status, editors require verified access links, documentation resolving whether the official rollout occurred on July 30 or August 1, and technical evidence demonstrating the active availability and open-source distribution of the H3 model.

DeepSeek-V4-Flash: # MiniMax Launches H3, an Open Omni-Modal Model That Generates Video With Native Audio **The Chinese AI startup claims H3 can jointly understand text, images, video and audio, and produce clips up to 15 seconds long at 2K resolution.** **By RuntimeWire Staff** MiniMax has officially launched its H3 model, which the company describes as a "general-purpose omni-modal generation model" capable of jointly understanding multimodal contexts spanning text, images, video, and audio, according to a July 30 announcement from MiniMax Research. The company further promoted the release in an Aug. 1 post on X, calling H3 "an open model breaking the boundaries between tasks and modalities." The July 30 and Aug. 1 publication dates appear consistent: the research team published its technical announcement on July 30 at 7:00 PM CT, while the official corporate account amplified the launch on Aug. 1 at 12:50 AM CT. No contradiction is evident. MiniMax claims H3 can generate video with "native stereo audio" at resolutions up to 2K and lengths up to 15 seconds. The company has not specified which generation tasks are supported for text, images, or audio independently. It also has not disclosed training data, model architecture details beyond the "omni-modal" descriptor, or performance benchmarks. The model's "open" status is asserted but not defined in the supplied materials. MiniMax's announcements do not specify whether H3 is released under a permissive license, what parameters are available, or where the model weights can be accessed. The tweet's link points to a URL that is not fully resolvable from the assignment materials. **Verification needed before publication:** The reporter should confirm H3's actual availability on platforms such as Hugging Face or GitHub, and clarify whether "open model" means open weights, open source, or another definition. Without this verification, a story asserting the model's open status cannot be responsibly published. The link in MiniMax's tweet should also be tested to ensure it leads to a live, substantive page.

Google: Gemini 3.7 Flash: **DeepGrove Releases Maple-Preview Ternary MoE Model for On-Device Inference** *The open-source preview posts high reported speeds on consumer hardware, though the developers note gaps on agentic tasks.* AI developer DeepGrove has unveiled Maple-Preview, an open-source 20B-A1B reasoning model designed to run directly on consumer hardware such as laptops and smartphones. The model features a 5.31 GB checkpoint and supports a context window of up to 131,072 tokens. According to DeepGrove, the system is natively trained as a ternary-weight network rather than post-quantized from full precision, an approach the company argues replaces matrix multiplications with additions to reduce arithmetic workloads. Its architecture features 24 layers and 256 experts in a mixture-of-experts (MoE) configuration, using hybrid sliding-window and global attention to limit KV-cache memory expansion. DeepGrove reported substantial generation speeds in its internal testing, claiming the preview achieved 218 tokens per second on an Apple M4 Mac mini and 127 tokens per second on an iPhone. The company asserted that Maple-Preview runs 5 to 16 times faster on its test setups than models including Gemma 4, Qwen3.5, and gpt-oss. DeepGrove also reported that the model solved Problem 1 of the 2024 International Mathematical Olympiad (7/7) during a test on an M5 Pro MacBook Pro. However, DeepGrove framed the release as an early preview focused strictly on raw reasoning capabilities rather than general-purpose deployment. The company explicitly acknowledged that the model may underperform on agentic benchmarks in its current state, adding that it plans to conduct extended training to improve overall performance ahead of a full release.

DeepSeek-V4-Flash: # DeepGrove Launches Maple-Preview, an Open-Source 20B Ternary Model Claiming 120+ Tok/s on iPhone **On-device reasoning model uses natively trained ternary weights to challenge assumptions about low-precision AI performance, but company cautions it's a preview focused on raw reasoning.** SAN FRANCISCO — DeepGrove today released Maple-Preview, an open-source 20B-A1B ternary-weight reasoning model that the company reported achieves 127 tokens per second on an iPhone — a speed it claims is 13× faster than the 1-bit Bonsai 27B model on the same device when generating a response to a common prompt. According to a company post on Hacker News, the model's checkpoint is 5.31 GB and supports a 131,072-token context. DeepGrove reported that Maple-Preview runs at 218 tokens per second on an M4 Mac mini and at 281.5 tokens per second on a MacBook Pro with an M5 Pro chip when solving IMO 2024 Problem 1. DeepGrove stated that Maple-Preview is "SOTA in its weight class" and "even competitive with larger models," though it characterized the release as a preview. The company noted the model "may underperform on agentic benchmarks" and said it intends to "continue improving general performance through extended training before Maple's full release." The company positions Maple-Preview as a native ternary-weight network — meaning it was trained at low precision from the start, rather than converted from a full-precision model — which DeepGrove called "fundamentally the wrong approach" for optimizing both performance and efficiency. "At ultra-low bitwidths, matrix multiplication can be effectively replaced with additions," the company wrote, envisioning "a world where most inference is done on smaller, personalized models." The architecture uses a 24-layer, 256-expert configuration with hybrid sliding-window and global attention, which DeepGrove said was chosen to bound KV-cache growth after testing configurations on a Mac mini for inference speed. Maple-Preview is available for chat at chat.deepgrove.ai.

Google: Gemini 3.7 Flash: **Latitude Health Raises $2M Toward $3M Offering for Utilization Management Platform** *An SEC Form D filing reveals two investors have backed the San Francisco-based startup in its first concrete funding disclosure.* San Francisco-based healthcare technology startup Latitude Health, Inc. has secured $2 million in funding, according to a Form D filed with the U.S. Securities and Exchange Commission on August 7, 2026. The regulatory filing indicates that the Delaware corporation has established a total offering size of $3 million, leaving $1 million remaining to be sold. A total of two investors have participated in the offering as of the filing date. This filing represents the first concrete financing disclosure for the company, which appears to have emerged from stealth around early 2025. According to the SEC document, Latitude Health is led by Charles Feerick, who is listed as an executive officer and director. Jarred Bressner and Chris Palmieri also serve as directors for the company. The Form D filing does not disclose the specific identities of the participating investors, nor does it characterize the terms or nature of the securities sold beyond the reported financial figures. Latitude Health operates an AI-native platform designed for utilization management (UM) and prior authorization workflows in healthcare. According to marketing statements published by the company, the platform is designed to automate 75% of manual administrative work, double clinical review throughput, and reduce clinician burnout. These operational metrics and workflow automation claims are self-reported statements by the company and have not been independently tested, verified, or evaluated by RuntimeWire.

DeepSeek-V4-Flash: # Latitude Health Discloses $2M Raise for AI-Powered Utilization Management Platform **SAN FRANCISCO** — Latitude Health, Inc. has disclosed $2 million in new financing through a securities offering, according to a Form D filed with the SEC on August 7, 2026, marking the first concrete funding disclosure for the stealth-stage healthcare technology company. The Delaware corporation, headquartered in San Francisco, sold $2,000,000 toward a total $3,000,000 offering. The filing indicates only two investors have participated to date. The securities were sold on August 7, the same day the filing was submitted. According to the filing, company leadership includes Charles Feerick as Executive Officer and Director, with Jarred Bressner and Chris Palmieri also serving on the board. Latitude Health describes itself as operating an "AI-native platform for utilization management and prior authorization." The company claims its technology can automate 75% of manual work, double review throughput, and reduce clinician burnout. These performance claims are company assertions and have not been independently verified. The company appears to have emerged from stealth around early 2025 and maintains a web presence at latitudehealth.com. The Form D filing does not identify specific investors or characterize the securities beyond establishing that they were sold in this offering. The total offering size suggests the company may seek an additional $1 million from future investors. **Editor's note:** This story is based solely on Latitude Health's SEC Form D filing. While the filing provides verifiable details about the offering amount, structure, and company leadership, the company's product claims regarding automation rates and clinical outcomes are reported as stated by Latitude Health and have not been independently confirmed. Further verification would require interviews with company executives, customer testimonials, or third-party validation of the platform's performance metrics.

Google: Gemini 3.7 Flash: **Headline:** GitHub Repository Claims High-Throughput DeepSeek V4 Flash Stack on AMD MI300X **Dek:** An uncorroborated repository reports up to 168.6 tokens per second on single-GPU hardware, but available evidence remains insufficient for independent verification. A newly published GitHub repository claims to provide a production-validated software stack capable of running DeepSeek’s `deepseek-ai/DeepSeek-V4-Flash-0731` model on a single AMD MI300X (gfx942) accelerator. Published on August 4, 2026, by user `ryanzhou`, the repository `deepseek-v4-flash-mi300x` contains a digest-pinned vLLM ROCm compose deployment, ten byte-identical correctness and performance overlays with SHA-256 pins, reference unified diffs against upstream vLLM, ROCm-triton, and AITER, and AITER GEMM tuning tables. According to commit metadata in the repository, the setup achieved measured results of 168.6 tokens per second median single-stream decode and between 7.9K and 8.5K tokens per second in tuned prefill. The claims touch on ongoing industry efforts to optimize software for AMD's MI300X accelerator, which launched in December 2023 featuring 192GB of HBM3 memory compared to the 80GB found on Nvidia’s competing H100. However, *RuntimeWire* assesses that the available evidence is currently insufficient to publish these claims as established facts. The performance figures and production claims originate entirely from a single, newly created repository with one commit, zero stars, and zero forks. The provided sources contain no third-party verification, independent benchmarks, or external confirmation of deployment validity, and no details are provided regarding the author's professional affiliation or background. To support a fully publishable news report, *RuntimeWire* would require additional verification, including: - Independent reproduction of the prefill and decode benchmarks on physical MI300X hardware. - Technical review and testing of the ten patch overlays against upstream vLLM, ROCm-triton, and AITER codebases. - Third-party corroboration confirming that the software configuration operates reliably in production environments.

DeepSeek-V4-Flash: # GitHub Repo Claims DeepSeek V4 Flash Achieves 168 Tokens/Sec on Single AMD MI300X **A newly published repository purports to deliver production-grade inference for DeepSeek's latest model on AMD hardware, but key claims remain unverified.** **SAN FRANCISCO** – A GitHub repository published early Tuesday claims to enable running DeepSeek V4 Flash on a single AMD MI300X accelerator, reporting median single-stream decode speeds of 168.6 tokens per second and prefill throughput between 7,900 and 8,500 tokens per second. The repository, created by user "ryanzhou" under the name `deepseek-v4-flash-mi300x`, describes itself as a "production-validated stack" for the model `deepseek-ai/DeepSeek-V4-Flash-0731`. It includes code patches, tuning tables, and a README documenting measured results. According to the repository's commit message, the stack uses "digest-pinned vLLM ROCm compose deployment" with "ten byte-identical correctness/performance overlays with SHA-256 pins." The project also provides "reference unified diffs against upstream vLLM/ROCm-triton/AITER" and "AITER GEMM tuning tables." The README – referenced in the repository's summary – reports three key performance figures: "168.6 tok/s median single-stream decode" and "7.9-8.5K tok/s tuned prefill." However, the repository does not specify the methodology, test duration, hardware configuration (beyond a single MI300X), or software environment used to obtain these results. These figures are reported results from the repository author and have not been independently verified. RuntimeWire could not confirm whether the patches successfully build or execute, whether the claimed performance is reproducible, or whether the repository has been reviewed by AMD, DeepSeek, or any third party. The repository contains zero stars and zero forks at time of publication, suggesting minimal community validation. The single commit was made approximately 22 minutes before the story was filed. Additional context is available from a June 2026 blog post by Doubleword, an inference cloud provider, which documented challenges bringing DeepSeek-V4-Flash to AMD MI300X hardware. That post, which is not affiliated with the new repository, described the MI300X as "underappreciated" due to software limitations despite offering 192GB of HBM3 memory at roughly half the list price of NVIDIA's H100. **Assessment:** The available evidence is insufficient for publication as a standalone news story. The repository's claims are unverified, lack methodological detail, and have received no community validation. Additional verification would require: (1) independent reproduction of results by a third party; (2) confirmation of the author's identity and technical credentials; (3) review of benchmark methodology including batch sizes, precision, and prompt lengths; and (4) confirmation that the model referenced (`DeepSeek-V4-Flash-0731`) is an officially released model from DeepSeek.

Google: Gemini 3.7 Flash: **AGent Energy Raises $11M Seed Round to Tap Behind-the-Meter Generation** *The Houston startup has secured $17 million in total funding to scale its team and target an estimated 200 GW of backup power across commercial and institutional sites.* Houston-based AGent Energy announced Thursday that it has closed an $11 million Series Seed financing round, bringing the startup’s total funding to $17 million across its first 12 months of operations. The financing was co-led by Spero Ventures and MassMutual Ventures, with participation from Intrepid Investment Management alongside existing investors CIV and Zero Infinity Partners (ZIP). The round follows an initial $6 million financing closed shortly after the company's founding. In connection with the investment, Stephen Wemple of Spero Ventures will join AGent’s board of directors, and Aram Ouligian of MassMutual Ventures will join as a board observer. AGent stated that it plans to deploy the fresh capital to expand its workforce and accelerate efforts to tap an estimated 200-plus gigawatts of behind-the-meter generation across commercial and industrial facilities, as well as the MUSH sector—comprising municipalities, universities, schools, and hospitals. The company framed the 200-plus GW figure as an industrywide estimate of idle on-site generation that it aims to target, rather than capacity currently under its operational control. The company pitches its AI-based software platform as a response to mounting power grid stress, pointing to surging data center load growth and supply constraints, such as PJM’s recent capacity auction clearing at its price cap. According to AGent, its platform aggregates, orchestrates, and monetizes idle on-site backup generation, converting existing equipment into rapidly dispatchable distributed power plants at no cost to asset owners. AGent also claims it is actively participating in three major North American wholesale power markets, asserting that its platform has successfully delivered capacity during grid emergency events across the PJM, MISO, and ERCOT footprints.

DeepSeek-V4-Flash: **AGent Energy Secures $11M Series Seed to Tap 200+ GW of Idle Behind-the-Meter Generation** **The AI-driven distributed power plant developer, which says it has raised $17 million in total funding in its first year, plans to scale its team and expand market operations across commercial, industrial and institutional sites.** HOUSTON – AGent Energy announced Wednesday it has closed an $11 million Series Seed financing co-led by Spero Ventures and MassMutual Ventures, with participation from Intrepid Investment Management and existing investors CIV and Zero Infinity Partners (ZIP). The company stated the round brings its total funding to $17 million within 12 months of founding, following an earlier $6 million financing from CIV and ZIP. AGent describes itself as a developer of AI-driven distributed power plants. The company says its platform aggregates, orchestrates and monetizes backup generation already located at commercial, industrial, and mission-critical facilities, including AI data centers. According to the company, the equipment is already built and idle for most of the year. AGent states it is currently dispatching capacity in PJM, MISO, and ERCOT, having responded to grid emergency events in those markets. The company said it will use the capital to scale its team and accelerate efforts to unlock what it estimates as 200+ gigawatts of behind-the-meter generation. Quoting the CEO and Co-Founder, the company said it targets the commercial and industrial sectors as well as the MUSH sector (municipalities, universities, schools, and hospitals). Company representatives attributed the need for such capacity to strained grid conditions. AGent pointed to data center load growth and PJM’s most recent capacity auction, which the CEO described as reflecting “the urgency of the grid reliability challenge.” The company has not disclosed current capacity under management, customer counts, or revenue figures. In connection with the financing, Stephen Wemple of Spero Ventures will join AGent’s board of directors, and Aram Ouligian of MassMutual Ventures will join as a board observer. *(This story attributes all claims about market conditions, generation capacity targets, and dispatch events directly to AGent Energy or its representatives. It does not independently verify figures related to PJM auction results, data center load projections, or the company’s stated 200+ GW estimate.)*