Entries for August 14, 2026

AS WE SAID ~Opus 4.6 at home. 42.2 on DeepSWE puts this tiny model somewhere near GLM 5.2 on coding related tasks. This is nuts! Obviously there are benefits to larger models that you will notice compared a 27B model. Some of you are so pedantic. You completely miss the point. The point is that for many tasks, especially tasks that are useful for doing real work, you can get REALLY close. At home. On your own hardware. That is amazing!!! Will you use this for the most difficult engineering tasks? No of course not. But guess what? You're not using Opus 4.6 either! For any task that you would delegate to such a model, this model will also do great. I received a ton of comments about being skeptical about how good this would be, but this was the only logical result. Qwen 3.6 27B was still among the best small models, any improvement would have made this SOTA for its size. So this should be no surprise, it's SOTA for its size. :) Sorry for the victory lap, but i'm just excited. This is great news for everyone.

Part 2! A part 3 is coming as well, seriously the website generation is so close to 3.8 max, it's insane we have this in a local model!

holy shit! look at the table qwen just published for the 27b. beating opus 4.6 max on computer use, 84.3 vs 72.7 on osworld. beating it on mobile use, 81.9 vs 62. beating it on multimodal software engineering. and visual math isn't even close, 94.6 vs 65.5. and i'll verify what i can locally, but the shape is unmistakable. this thing runs on a single rtx 3090. a 2020 gamer card. native vision, 262k context, apache 2.0, and the gguf ladder went live the same minute as the weights. frontier table numbers in a size that fits 24gb of vram, this just added five more years to every consumer card in the world. q4_k_m is already flying onto mine. i will drop honest numbers tonight.

@Alibaba_Qwen Thank you thank you Qwen team! Qwen3.8-27B is just amazing! 🥰 We made Dynamic GGUFs so you can run it locally: huggingface.co/unsloth/Qwen3.… The model is now also supported for running and fine-tuning in Unsloth Desktop.

We promised open weights for Qwen3.8. Now, time to meet them! 🎉 ⚡ Qwen3.8-27B: - A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows. - 262K native context, easily extendable to 1M tokens via YaRN. - Built for builders. Highly efficient, high-quality, and licensed under Apache 2.0. 🚀 The open weights for Qwen3.8-2.4T-A95B (Max-level) have also been released recently. Whether you're shipping lightweight applications with Qwen3.8-27B locally or building agents with Qwen3.8-2.4T-A95B, they're yours now! Download, deploy, and build something we haven't imagined yet. 👀👇 - Hugging Face： huggingface.co/collections/Qwen/qwen38 - ModelScope：

DeepSeek Harness implements RSI (Recursive Self-Improvement) for agents. The paper provides the missing infrastructure layer that makes continuous, safe self-modification practical rather than theoretical. Core problem it solves for RSI: Most RSI discussions focus on an agent improving its model weights, code, or algorithms. Real agents, however, also need to continuously rewrite their own harness — the runtime system that supplies tools, memory, sandboxes, orchestration loops, skills, sessions, and interfaces. Without proper guarantees, a self-modification can: Leave irreversible side effects Break dependency chains Destroy the recovery mechanism itself Force a full process restart (interrupting continuous evolution) This is the classic “agent kills itself while trying to improve itself” failure mode. The proposed solution: Spatiotemporal Composability The paper formalizes two orthogonal properties required for safe RSI: Temporal composability (time dimension): Every side effect a component produces must carry an explicit inverse. When the agent unloads or replaces a component, the runtime can automatically and completely undo all its effects (state changes, resource registrations, listeners, etc.). Bad self-improvements become reversible. Spatial composability (space dimension): Components declare their dependencies. When a provider appears, disappears, or is replaced, the system reactively recomputes only the affected parts of the dependency graph. The agent does not need manual boot-order management or global restarts. These are unified into a single context programming paradigm and given a formal calculus whose metatheory proves that the properties lift from individual components to an entire interleaved system. Key theoretical result relevant to RSI Under reasonable conditions (independence + acyclic dependencies), the system has confluence: after any sequence of dynamic loads, unloads, and replacements, the final stable state is equivalent to the state that would result from assembling the final configuration from scratch. Intermediate evolutionary steps leave no permanent traces. This is crucial for open-ended self-improvement — the agent can explore many self-modifications without accumulating irreversible technical debt. Practical implication Implemented in the Cordis meta-framework (used by DeepSeek Harness), this turns the entire agent stack into hot-swappable plugins. An RSI-capable agent can therefore: Generate, deploy, and replace its own tools, memory systems, orchestrators, etc., while remaining online Isolate failures locally Roll back failed improvements cleanly Keep evolving indefinitely without process restarts In short: the paper shifts the RSI bottleneck from “can the model write better code?” to “can the agent safely rewrite the substrate it is running on?” It supplies the formal and engineering foundation for the latter. github.com/cordiverse/paper