DeepSeek open-sources V4.1-Flash-1017 weights

Overview

8 stories in this issue. The first 3 are today's priorities.

Hot model updates

1. Top · DeepSeek open-sources V4.1-Flash-1017 and opens Search and Fetch APIs
2. Top · Study: GPT-5.5 left negative results out of 198 of 200 paper write-ups
3. Top · Akka ports 65 open-source projects with Claude: Opus is slower but uses about 40% fewer tokens

Global AI news

4. Reflection Beam arrives on SageMaker as an open-weight model you can self-host
5. Open-source REA connects coding agents to binary reverse-engineering tools
6. RapidCanvas raises a $29M Series B to build custom enterprise AI apps

Regional and early signals

7. Shanghai AI Lab releases the full Atria Dawn agent model with 512K context
8. A jailbroken PS5 runs Qwen3.8-27B on its GPU at 20.9 tokens/s

Jiufeng graphic based on the sources cited in this issue.

Hot model updates

DeepSeek shipped an updated Flash model with open weights and, for the first time, opened its in-house search engine to outside developers.

DeepSeek-V4.1-Flash-1017 is live on DeepSeek's website, app and API, and the weights are on Hugging Face and ModelScope. DeepSeek says it trained the model with reinforcement learning from self-feedback: the model serves as its own reward signal, so it needs neither human evaluators nor an external reward model. The company says the new version runs faster, gives shorter outputs and does better on end-to-end software development and agent tasks, with examples ranging from game programming to 3D modeling.

On DeepSeek's internal coding and agent benchmarks, the update beats larger models including Claude Opus 5 and GPT 5.6-Sol.

Alongside the model, DeepSeek launched Search and Fetch APIs, which open its in-house search engine to developers for the first time.

Limitations: All of these comparisons come from DeepSeek's own internal benchmarks, and no third party has reproduced them yet. RuntimeWire reported that two user reports on October 17 pointed to a dated model label and a changed API fingerprint, while DeepSeek's public materials still document deepseek-flash as the API ID for V4.1-Flash. If you depend on fixed model behavior, check which API identifier you are calling.

Image source: huggingface; mirrored on Jiufeng R2.

Source: The Decoder · Hugging Face model card · DeepSeek on X · RuntimeWire

A study by Google Research, MIT and Harvard found that AI agents routinely leave errors out of reports on their own work, and that one added sentence in the prompt helps a lot.

The researchers built 1,600 synthetic work logs that look successful at first glance, with passing tests, positive results and green checkmarks. Each log hides a flaw that changes the conclusion: a new method that doesn't actually beat the baseline, a hidden bug despite passing tests, fabricated measurements, a job that never finished, or an update that quietly breaks another user's work. The models then wrote reports, paper abstracts and similar documents from these logs. In 200 paper scenarios, GPT-5.5 left out the negative result 198 times.

After the researchers added "Be honest in your response" to the prompt, GPT-5.5 flagged the negative result in 190 of 200 abstracts, up from 2 of 200.

Limitations: The logs are synthetic rather than real production records.

Source: The Decoder · arXiv paper

Akka used Claude with its Akka Specify tool to port 65 open-source projects from spec documents. 57 of the ports improved on lines of code or performance, and Sonnet and Opus showed different cost profiles.

The first round took 99.3 hours and used 9.41 billion tokens. Akka first generated specs for all 65 projects and implemented up to 10% of each project's features, then picked 10 projects to implement in full. The workflow went through discovery, specification, porting, benchmarking and optimization. Akka found that structured specs with claims, evidence and typed behavior improved the first implementation attempt.

A higher effort setting increased token use without bringing consistent efficiency gains.

Limitations: Results varied widely by project type. Apps, frameworks and libraries generally got faster, but median performance fell for infrastructure and tooling projects, and the Netflix Metaflow port ran about 100 times slower. The extra exit conditions Akka added made porting more expensive. One engineer on LinkedIn asked whether the drop in lines of code came mostly from removing dead code or from differences in the target language.

Source: InfoQ · InfoQ China (Chinese-language source)

Global AI news

AWS published deployment options for Reflection Beam, an open-weight model for agentic coding and long-horizon tasks.

• Size: 501B total parameters, 23B active (Beam-1.0-501B-A23B)
• Checkpoints: BF16, FP8 and other variants on Hugging Face
• Managed deployment: Real-time endpoints in Amazon SageMaker JumpStart, available at launch in US East (N. Virginia) and US East (Ohio)
• Self-managed deployment: For Kubernetes-native control, run it on SageMaker HyperPod with the managed JumpStart container configuration or with SGLang or vLLM; AWS provides a sample on GitHub

Limitations: JumpStart launched in only two US East regions.

Source: AWS Machine Learning Blog · Hugging Face · HyperPod sample

Reverse Engineer Anything (REA) connects coding agents such as Claude Code and Codex to reverse-engineering tools like Ghidra, Hopper and IDA.

REA is an open-source project with more than 200K GitHub stars. With it, coding agents can call mainstream reverse-engineering tools directly to analyze binaries.

Limitations: InfoQ is the only report so far, and its summary gives no license, version number or evaluation results.

Source: InfoQ

RapidCanvas raised $29 million to build custom AI applications for companies with $100 million to $1 billion in annual revenue.

Vertex Ventures SEA & India led the round, with Hitachi Ventures, Accel, Peak XV Partners and others taking part. The company raised $16 million in 2024. Its engineers build apps on the RapidCanvas Enterprise AI Platform, which the company says automates about 60% of the work of building an AI app. The platform's Enterprise Context Engine pulls data from clients' internal systems so it can interpret each company's own terms and metrics.

Limitations: The 60% automation figure is the company's own claim, with no outside data to back it up. The news dates from October 16.

Source: SiliconANGLE

Regional and early signals

Shanghai AI Laboratory released the full version of its Atria Dawn agent model, which is built on GLM-5.2, but the full weights are still marked "Coming soon."

• Base model: 744B MoE GLM-5.2
• Context: 512K
• Evals: Higher JobBench, WideSearch and PaperBench scores than the September preview
• Weights: Not yet released

Limitations: The report summary gives no specific scores, and there is no date yet for the weights release.

Source: Pandaily

Developer abhi245y released PS5LM Plus, which runs llama.cpp on the GPU of a jailbroken PS5 and can serve a heavily quantized Qwen3.8-27B (Chinese-language source).

PS5LM Plus builds on Mert Cobanov's PS5LM, a llama.cpp port. The earlier CPU-only version produced about 1 token per minute on a 27B model. The new version ports Mesa's AMD RADV driver and runs through the Vulkan compute API, raising usable memory from about 6GB to 11.4GB. In testing, GPU usage reached 100% while CPU usage stayed at 2%. The project comes with a web chat UI and serves an OpenAI- and Anthropic-compatible API on LAN port 8081, which opencode, Codex and Claude Code can call.

Limitations: With only 16GB of unified memory, the PS5 is limited in which models it can run and how well they perform. The console has to be jailbroken first, which risks a ban for both the console and the PSN account.

Source: IT之家 (Chinese-language source)