Best Open LLM

research.ibm.com

How llm-d makes the most of the hardware you already have

IBM Research and Red Hat deployed a 753B open model on H100 GPUs, serving thousands of concurrent coding agents at 5-10x lower cost than commercial APIs.

igeek.gamer-geek-news.com

Original post on igeek.gamer-geek-news.com

www.frenchtechjournal.com

Mistral AI Raises €3 Billion at €21 Billion Valuation to Keep Europe in the AI Race

Mistral would like to remind everyone that reports of it being "cooked" are premature. The Paris company said Tuesday that it has raised €3 billion at a post-money valuation of more than €21 billion (roughly $24 billion), in what it bills as the largest equity round ever completed by a privately held European technology company. The funding comes three years after its launch. For a startup that began as a small group of ex-DeepMind and ex-Meta researchers with a strong opinion about open models, that is a decent run by any measure. Unless you compare it to private labs in the U.S., in which case it still falls short. Anthropic has raised around $100 billion so far this year ahead of an IPO that could value it at as much as $2 trillion. Reuters pegs OpenAI at $852 billion as it heads to a public listing this year. So, yeah, against that, €3 billion kinda feels like a rounding error. And yet... Seen from inside Europe, the numbers look different. Mistral becomes the continent's second-highest-valued private tech group and its fastest-growing company. The valuation has nearly doubled in a year, from around €12 billion. As for French milestones, the funding is remarkable on several levels. On The French Tech Journal funding tracker, the round puts the ecosystem past €8 billion raised so far in 2026, already topping €7.8 billion for 2025. For vets of the French Tech scene, it was just back in 2017 that €3.1 billion represented the **_total funding for the entire year._** Chart courtesy of Dealroom For the latest round, Samsung Electronics led the funding while the EU-backed Scaleup Europe Fund, managed by EQT, and existing shareholder PSG Equity came in as co-leads. Advent, funds managed by BlackRock and the Grand Duchy of Luxembourg, joined as new investors. Almost everyone already on the cap table re-upped: a16z, ASML, Bpifrance, BNP Paribas CIB, DST Global, General Catalyst, Index Ventures, Lightspeed, Nvidia, Salesforce Ventures. Samsung isn't just a tourist. In an interview with the Financial Times, Mistral AI CEO Arthur Mensch called the investment "another validation" of Mistral's usefulness in high-end manufacturing, following ASML's €1.3 billion cheque last year. Samsung plans to use Mistral's AI to improve its chip manufacturing systems. Although Microsoft agreed in July to spend billions of dollars on Mistral's computing infrastructure in Europe, the Redmond giant apparently did not participate, Mistral AI chief financial officer Johan Bergqvist told Reuters. ## What the money buys Despite a general freakout last month after the announcement that Mistral AI would build compute primarily focused on inference, the company stressed in the funding announcement that: "The round will significantly expand Mistral's **_frontier research_** , which is the foundation underpinning its infrastructure, products and sovereignty." In the FT, Mensch said the company believes it will now have the compute to match the strategy used by Chinese AI labs, which allowed them to match the frontier training of OpenAI and Anthropic at a fraction of the cost. “With this fundraising we will be controlling an amount of compute that is very comparable to what the Chinese labs have,” he said. “We have built Mistral to train models and to scale them. We will continue to do that, and that’s really the purpose of this fundraising.” Mensch told CNBC the money goes toward more infrastructure, including Mistral's own data centers, and that the company will train "bigger and faster models." He said Mistral is on track for $1 billion in annual recurring revenue by year-end, with growth strongest in Asia and North America. Mensch went further with CNBC, saying he expects "to be beating" that figure "if everything happens as they are trending." ## Sovereignty, with an asterisk Mistral AI continues to place sovereignty at the core of its identity and pitch. "The fact that the EU or Europe has to have its own kind of AI provider in the game is important," Bergqvist told Reuters. "We have seen in the past that there is politics involved in the access to these solutions, and it's important that you make sure that you maintain access and do not compromise your kind of supply chain." Mistral's answer is the full stack: open-weight models, the infrastructure they run on, and the products that put them into production. Control means data that stays inside an organization's walls, customizable models, private compute, and auditable systems. The company now works across 20 countries with more than 125 enterprises, including Airbus, ASML, and HSBC. Still, eyebrows were raised in August when Mistral said it would commercialize GLM-5.2, a model built by the Chinese firm Z.ai. Mensch defended it to the FT as giving customers choice while still offering "regional controls" to protect their data. But he also hinted that it may just be a stopgap solution. Mensch told CNBC the models Mistral releases "very soon" will be "very competitive."

techyon.pages.dev

Running Qwen 3.8 27B at Q4 on 16GB VRAM at 200K CTX at 50t/s

A user successfully runs the Qwen 3.8 27B model at Q4 precision on a 16GB RTX 5070Ti, achieving 50 tokens per second with a 200K context window. The setup uses Unsloth's UD-IQ4_XS GGUF quantization an

arxiv.org

Multi-Step Tool-Calling over Korean Open Public APIs: A Benchmark and a Data-Synthesis Recipe

Data-sovereignty regulations increasingly require public institutions to deploy open-source, on-premise LLM agents that chain multiple tool-calls across live government APIs. However, open-source models consistently underperform in this multi-step setting, and no existing benchmark measures the gap…

dlvr.it

Benchmarking small LLM inference on SageMaker AI: G7 vs G5 and G6

Benchmark two 30B Mixture-of-Experts models, Qwen3-Coder-30B and NVIDIA Nemotron-3-Nano-30B, across G5, G6, G6e, and G7 GPU instances on Amazon SageMaker AI. Compare throughput, latency, and cost-per-token, and see how G7's NVIDIA Blackwell GPUs deliver measurable price-performance gains for real-time LLM inference.

gezel.com

1.26251 — 8 September 2026

Engines that reserve memory instead of colliding, llama.cpp v0.4 loading controls, GLM 5.3 on DwarfStar, and catalogs built from a folder of Markdown.

www.tomshardware.com

Benchmarking Qwen 3.8 27B on RTX 5090 and beyond — VRAM capacity alone can't overcome severe software and inference engine bottlenecks

Following the release of Qwen 3.8 27B, we put our trusty hardware to the test to see which hardware might be best suited for running this open-weight AI model.

whatsnew.fyi

open-design open-design-v0.22.0

🔎 66 PRs · 10 contributors · 7 days — We evaluated over a dozen models using the OpenDesign Harness to compare design quality and cost. DeepSeek V4.1 Flash achieved 98% of top-ranked GPT-6 Astra’s average score at just 1% of its average cost. Explore OpenDesign Arena to find the right model for…

unsloth.ai

GLM-5.3-Flash: How to Run Locally | Unsloth Documentation

Run the new GLM-5.3-Flash aka ox-alpha model by Z.ai.

dev.to

Leanstral 1.5 is a what..?

I was doing the do out on the internet, and you do, and came across Mistral AI Leanstral 1.5. Which calls itself: > An updated Lean 4 formal proof engineering model optimised for automated theorem proving and autoformalization. 119B total parameters, 6.5B active. I was like, wow, thats a thingymabob of the first order. The last time someone said "formal methods" to me was in 1998, when a bloke called in a computer lab in Surrey. I had not a scooby doo what that was, and here we are more than a quarter of a century later, and solving something cool with formal methods is still on my bucket list. Boo hoo. I had to check the blog to ask what-the-actual-fluffs this is all about Leanstral 1.5: Proof Abundance for All. They are letting you use it for free. So you can go wild having it formally prove you don't have bugs in your mission-critical systems code! Yipee! Well, it turns out I need a formal proof that my Unbounded Viewstamped Replication Revisited code doesn't have many problems. So I thought I'd ask Fable 5.1, GPT 5 and GLM-5.3-Flash to kick the tyres on getting the Leanstral model to write out formal proofs in pretty Greek characters and debug it. Huh you did what!? Well, lets say a model was not good at, I dunno, coding in a rare language, like Felienne Hermans’ Hedy. You could fine-tune a model on that and make a language server that lints the code so that a special model trained on a fast set of examples can happily write code in Arabic and right-to-left with all the different and rare Arabic numerals. That would be so darn cool. Well, Mistal AI have done that for the LeanProver language Lean. They have a VS Code Plugin. On their blog, they show Claude can find real solutions to real problems in that system, but at a very high cost. I told Claude to call Leanstral if it needed help in writing Lean proofs that could be solved. Which it duly did when it got a bit stuck. Here is what it wrote: So there you go! So, like, you know you could, I dunno,

soundcloud.com

zai-org/GLM-5.3-Flash

We introduce GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series. With 320B total parameters and just 18B active parameters, it outperforms GLM-5.2 across benchmarks and real-world

z.ai

GLM-5.3: Frontier Coding with Emergent Cyber Capabilities

Z.ai's open-weights 753B MoE coding model: 31.4% on Z.ai Code Bench at ~50K output tokens vs Claude Opus 4.8's 29.5% at ~120K.

aichina.news

Meet V3OS-nym2: A Lightweight LoRA Adapter for FLUX on Huawei's Ascend Ecosystem