coding model

Meta

2026-09-02 20:19:15

Meta Releases Muse Spark 1.3 Model, Advances Personal AI Agent Development

Meta announced on Wednesday the release of Muse Spark 1.3, an update to its AI model that significantly improves performance in coding and agentic tasks. According to Alexandr Wang, Meta's AI head, the new model is "competitive with frontier models" and paves the way for future personal AI agents that can work on behalf of users around the clock. The pricing remains unchanged from the previous version, which Wang described as an "aggressive" strategy. Meta also highlighted the popularity of its "contributor tier" option, which allows the company to use developers' work to improve the model, thereby reducing coding product costs. A "meaningful double-digit percentage" of developers have opted in. Wang noted that safety has become a key internal issue as model capabilities improve, and Meta is increasing investments in safety and alignment. Muse Spark 1.3 will be available on Muse Code and Meta API on the same day, with the highest reasoning version to follow after additional safety testing. This update reflects Meta's ongoing commitment to advancing AI capabilities while maintaining a developer-friendly ecosystem and prioritizing safety as the technology evolves.

00

Microsoft

2026-10-07 17:21:43

Microsoft launches an AI coding model that runs on personal laptops

Microsoft (MSFT.O) has introduced an artificial intelligence coding model that can run on a personal laptop, according to ChainCatcher. The brief update only states the product launch and its ability to operate on consumer notebook hardware. It does not provide technical specifications, release timing, pricing, model size, or target users. No additional product details were included in the source note.

20

JetBrains

2026-10-09 08:06:01

JetBrains open-sources Mellum2.1 after posting a higher LiveCodeBench score than Qwen3.5-9B

JetBrains has released Mellum2.1, an open-source coding model positioned as an upgrade to Mellum2, which the company introduced in June this year. The new version keeps the same 12 billion total parameters and activates 2.5 billion parameters per generation, but JetBrains said it performs better at identifying code issues, editing files, and checking results. The company said it did not change the model architecture and instead improved performance mainly through reinforcement learning. According to JetBrains, the team ran millions of sandbox tasks across thousands of environments during training, with the model using terminal and file-editing tools inside real code repositories and receiving rewards when tests passed. In JetBrains’ own benchmarks, Mellum2.1 raised its success rate on SWE-bench Verified from 2% to 47%, slightly below Qwen3.5-9B’s 50%. On LiveCodeBench v6, however, Mellum2.1 scored 82%, ahead of Qwen3.5-9B’s 75.4%. JetBrains also said that under heavy-load inference on a single H200, the new model’s output throughput was close to twice that of Qwen3.5-9B. Model weights and a GGUF quantized version are now available on Hugging Face under an Apache 2.0 license for local deployment.

20

Reflection AI

2026-10-06 01:50:37

Reflection AI launches open-source coding model Beam, says reasoning matches GLM-5.2

AI startup Reflection AI has introduced Beam, an open-source coding model with 501 billion parameters. The company said Beam performs on par with Z.AI’s GLM-5.2 in advanced reasoning benchmark tests. Beam is built for coding and AI agent workloads and is now available for download from Reflection AI’s official website. Tech in Asia also noted that Reflection AI had previously received investment from NVIDIA. The release puts Beam into the open-source model field with a stated focus on software development and agent-based use cases, while Reflection is positioning the model around reasoning performance as a key comparison point.

40

Moonshot AI

2026-09-11 07:53:23

Kimi K2.8 Preview Rolls Out to Code CLI With 1M-Token Context

Moonshot AI has begun a full rollout of Kimi K2.8 Preview to Kimi Code CLI, according to BlockBeats. Users do not need to switch model IDs, as the existing kimi-for-coding endpoint will now point directly to the new model. Kimi said K2.8 Preview delivers overall performance close to K3 while running thinking tasks more efficiently. The model also adds the same thinking-intensity controls available on K3 and expands context length to 1 million tokens. By comparison, the previous K2.7 Code version supported a 256K context window and only allowed thinking to be turned on. The update effectively brings two core capabilities from K3 into Kimi’s default coding model.

860