THE 60-SECOND VERSION
● Qwen 3.8-27B scored 61.7 on SWE-Bench Pro. Claude Opus 4.6 Max scored 53.4. Apache 2.0, 262K context, runs on one 24GB GPU.
● tl;dv exposed 181,874 meeting records from 84,312 users across 35,003 domains. Reported January 28. Still open six months later.
● Google HEIR shipped as open source: compile a model, run inference on encrypted input, server never sees the data.
● 15 days left on Claude Sonnet 5 at 2 dollars per million input. September 1 it becomes 3 dollars.
Qwen 3.8-27B: the local model that beat a frontier model on real code
Alibaba released Qwen3.8-27B on August 14 under Apache 2.0. ModelScope timestamps the release at 15:00 UTC. It is a dense 27.78B-parameter multimodal model with a native 262,144-token context window, extensible to roughly 1M via YaRN.
The number that matters: 61.7 on SWE-Bench Pro. Claude Opus 4.6 Max scores 53.4 on the same benchmark. Qwen also takes QwenSWEBench (79.0 vs 63.8), LiveCodeBench v6 (90.3 vs 88.8), and IFBench instruction following (79.5 vs 62.5).
Against Meta Muse Glimmer 30B, the Apache 2.0 model released ten days earlier and covered here on August 11, Qwen wins every benchmark where both were run: SWE-Bench Pro 61.7 vs 53.4, QwenSWEBench 79.0 vs 63.8, IFBench 79.5 vs 77.0, GPQA Diamond 89.2 vs 83.5.
WHERE IT STILL LOSES
Opus 4.6 Max stays ahead on Terminal-Bench 2.1 (78.2 vs 73.0), repo-level generation on NL2Repo-Bench (47.6 vs 42.3), and the harder reasoning sets. Qwen is a better coding agent at 27B. It is not a better reasoner.
The honest caveat, and it is a real one: every launch number came from Qwen. Several benchmarks are in-house or modified. The SWE-Bench Pro comparison imports Anthropic's published Opus result rather than rerunning it under Qwen's harness. Independent replication is not in yet.
Hardware: official FP8 weights target server runtimes; Ollama lists 18GB packages. Community reports have it running on RTX 3090s and dual RTX 5080s. The 24GB floor is real but it is a floor. KV cache scales with context and concurrency, and a 27B model serving 262K requests to several users can double its footprint. Loading on your GPU and serving your workload on your GPU are different problems.
Hosted on OpenRouter at 0.45 dollars per million input, 3.20 dollars per million output.
We ran the full comparison against Muse Glimmer in a separate piece, including which one to pick for agentic coding versus document work.
tl;dv: 181,874 meetings, 84,312 users, six months of nothing
Security researcher bobdahacker disclosed that tl;dv, an AI meeting recorder used by more than 2 million people, shipped its Cloud Firestore meetings collection without tenant isolation. Any authenticated user, including a free-tier signup, could query every meeting record on the platform.
| What was exposed |
Scale |
| Meeting records |
181,874 |
| Unique users |
84,312 |
| Email domains |
35,003 |
| Government domains |
23 countries |
| Live recordings at any moment |
~1,000 |
Each record carried the creator email, conferencing provider, timestamps, recording status, and a conference ID. That last field is what turns a metadata leak into an access problem: the conference ID reaches the underlying Google Meet or Teams room. The researcher reports joining calls he was never invited to, including one hosted by the Malaysian Ministry of Education.
The disclosure timeline is the part that should worry anyone running a vendor review process. Reported January 28, 2026. Still unfixed in July. The CTO never replied. A separate internal employee-directory API was reachable with no authentication at all.
tl;dv holds SOC 2. That is the takeaway most people are drawing from this, and it is the correct one: a point-in-time attestation does not tell you whether a vendor answers the phone when a researcher finds a hole. SOC 2 Type II, evaluated over six to twelve months, is the minimum bar for anything recording your calls.
Google HEIR: inference on data the server cannot read
Google published HEIR, an MLIR-based open-source compiler toolchain that converts a pretrained model into a version that runs inference on homomorphically encrypted input. The server computes on ciphertext and never holds the plaintext.
Google demonstrated it on a deep learning recommender and on credit-card fraud detection. The developer entrypoint is a Python decorator: annotate which arguments are secret, and the compiler handles scheme selection and code generation against OpenFHE or Lattigo backends.
This is not a product you switch on tomorrow. FHE inference remains orders of magnitude slower than plaintext, and the practical workloads are narrow, structured models rather than a 27B transformer. What changed is that the compiler step no longer requires a cryptography team. For regulated sectors that currently cannot send data to a model at all, that is the blocking constraint being removed.
Deadlines still running
| Date |
What happens |
| Aug 17 |
OpenAI preliminary injunction response due in the Apple trade secrets suit |
| Aug 31 |
Claude Sonnet 5 goes 2 dollars to 3 dollars per million input, plus a tokenizer change adding 10 to 35 percent tokens on code |
| Aug 31 |
kimi-k2.5 and moonshot-v1 sunset, migrate to kimi-k3 |
| Oct 1 |
OpenAI vs Apple hearing |
| Oct 24 |
deepseek-chat and deepseek-reasoner deprecated |
What we are watching next
Independent SWE-Bench Pro replication on Qwen3.8-27B is the single most consequential open question in this batch. If the 61.7 holds under a third-party harness, the argument for paying frontier prices on routine coding agents gets significantly harder to make. If it drops five points, this is a good local model and nothing more.
The second thing to watch is whether tl;dv responds publicly. Six months of silence followed by a public disclosure usually forces a statement within a week. It has not come yet.
FAQ
Is Qwen 3.8-27B actually better than Claude Opus 4.6 Max?
On SWE-Bench Pro, QwenSWEBench, LiveCodeBench v6 and IFBench, per Qwen's own published numbers, yes. On Terminal-Bench 2.1, NL2Repo-Bench, GPQA Diamond and HLE, no. Treat it as a strong coding agent, not a general upgrade.
What GPU do I need to run Qwen 3.8-27B?
24GB is the practical floor for quantized weights, which puts an RTX 3090 or 4090 in range. Ollama packages are around 18GB. Long context and multiple concurrent users push memory well above the weight size because of KV cache.
Is my tl;dv data affected?
If your organisation used tl;dv before August 2026, assume meeting metadata, including creator email and conference IDs, was enumerable. Full transcripts were not in the exposed collection per the disclosure, but the conference IDs allowed joining live calls.
Does HEIR mean I can run ChatGPT on encrypted data?
No. HEIR targets structured models such as recommenders and fraud classifiers. Fully homomorphic inference on a frontier LLM is not practical at current performance levels.
When does Claude Sonnet 5 pricing change?
September 1, 2026. Input goes from 2 dollars to 3 dollars per million, output from 10 to 15 dollars. A tokenizer change lands at the same time that adds 10 to 35 percent more tokens on code, so the effective increase on coding workloads is larger than the sticker price suggests.