AI Dev Weekly #23: Qwen4 Architecture Preview, GPT-5.6 Lands in Kiro, Gemini Takes On Whisper
AI Dev Weekly is a Thursday series where I cover the week’s most important AI developer news, with my take as someone who actually uses these tools daily.
Four stories this week. Alibaba open-sourced an early look at the architecture it plans to use for Qwen4. OpenAI put the full GPT-5.6 family inside AWS’s Kiro coding agent. Google launched dedicated recorded and realtime transcription models. And GitHub moved Copilot’s cloud agent out of the IDE and into the Slack and Microsoft Teams conversations where work starts.
1. Qwen3.8-Flash-Next is really a public preview of Qwen4
Alibaba released Qwen3.8-Flash-Next on August 26 and opened its weights on Hugging Face and ModelScope. The name sounds like another incremental Flash release, but the more important detail is architectural: Qwen says this model previews the design that will underpin Qwen4.
The official Qwen release describes four areas of change:
- attention;
- residual connections;
- embeddings;
- training optimization.
It is a multimodal mixture-of-experts model and continues Qwen’s strategy of releasing a smaller architecture-preview model before rolling the design into the wider family. Qwen3-Next played that role before the Qwen3.5 through Qwen3.8 generations; Flash-Next is the equivalent early look at Qwen4.
Why this matters: Open weights turn an architecture announcement into something the community can inspect, serve, quantize, and benchmark. That is more useful than a closed preview with vendor charts alone. It also gives inference projects time to add support before the larger Qwen4 models arrive.
My take: Do not read “Qwen4 preview” as “Qwen4 is available.” This is an early architecture vehicle, not the finished flagship family. But it is still the most technically interesting release of the week because developers can examine the actual model rather than wait for a product announcement. If you deploy Qwen today, treat Flash-Next as an evaluation target, not an automatic replacement for Qwen 3.8 Max or a proven production model.
2. GPT-5.6 Sol, Terra, and Luna land in Kiro
OpenAI and AWS added the complete GPT-5.6 family to Kiro on August 24. Developers can now choose Sol, Terra, or Luna inside Kiro’s spec-driven workflow instead of treating Kiro as a model-specific coding environment.
The practical mapping is straightforward:
| Model | Best fit inside Kiro |
|---|---|
| GPT-5.6 Sol | Hard architecture, debugging, and long-running implementation work |
| GPT-5.6 Terra | Daily feature work where cost and capability both matter |
| GPT-5.6 Luna | High-volume routine changes, tests, and repetitive tasks |
OpenAI’s official Kiro announcement says Terra completed successful Terminal-Bench 2.1 tasks in Kiro at roughly 82% lower cost in joint testing. That is a vendor-reported environment-specific result, not proof that Terra is universally 82% cheaper than other models. Kiro’s structured requirements and technical designs are part of the measured setup.
The more durable change is model choice at each phase of a task. A team can use a lower-cost model for planning or mechanical implementation and reserve Sol for the places where deeper reasoning changes the outcome.
My take: This is less about another model picker and more about Kiro becoming a neutral orchestration layer. AWS letting OpenAI models compete inside its own coding agent is good for developers. The cost claim needs independent validation, but the ability to match model cost to task difficulty is immediately useful. See our GPT-5.6 in Kiro comparison for the detailed decision guide.
3. Gemini 3.5 Transcribe takes on Whisper with recorded and live APIs
Google now documents two dedicated speech-to-text models:
gemini-3.5-transcribefor recorded audio;gemini-3.5-transcribe-livefor realtime WebSocket sessions.
This is not just “send an audio file to a general Gemini model.” The transcription endpoints have their own feature matrix, limits, and pricing. Google’s recorded transcription documentation includes automatic language detection across 85+ languages, code switching, custom vocabulary, speaker diarization, and word timestamps.
There are important boundaries:
- recorded audio supports up to one hour, or 30 minutes when using diarization or word timestamps;
- the live endpoint uses ten-minute sessions;
- live transcription does not currently provide speaker diarization or word-level timestamps;
- smart formatting cannot be combined with diarization or word timestamps on recorded jobs.
Google estimates blended token costs at roughly $0.005 per recorded minute and $0.009 per live minute, though actual billing follows token usage rather than a guaranteed flat per-minute rate.
My take: Gemini wins when you want a managed API with built-in language switching, vocabulary hints, diarization, and a supported live path. Whisper still wins when audio must remain local, the application must work offline, or you want full control over deployment and retention. We published a full Gemini 3.5 Transcribe vs Whisper architecture comparison rather than pretending one option is universally better.
4. GitHub Copilot moves into Slack and Microsoft Teams
GitHub launched public previews that let teams start shared Copilot cloud-agent sessions from Slack or Microsoft Teams. Mention @GitHub in a supported conversation and the agent can investigate a problem, update issues, implement changes in a cloud sandbox, validate the result, and open a pull request.
The key change is not chat integration. It is shared agent work:
- the original conversation supplies context;
- teammates can inspect and redirect the session;
- the work continues asynchronously;
- the resulting pull request links back to the conversation;
- repository permissions still bound what the agent can access;
- administrators can require an extra approval for agent-authored pull requests.
The Slack preview is limited to Copilot Business and Enterprise organizations. The Teams preview is available with paid Copilot plans, consumes AI credits, and bills cloud sandbox usage separately.
My take: This is where coding agents were heading: the request starts in a stand-up or incident thread, not in an empty IDE chat. The risk is that an informal conversation becomes executable scope without enough review. Keep the pull request as the approval boundary, require extra approval for agent-authored work, and watch both AI-credit and sandbox budgets. For the broader product context, see our GitHub Copilot app guide.
Quick hits
- GLM-5.3-Flash weights shipped: Z.ai released a 320B-total, 18B-active multimodal MoE under MIT, with support documented for SGLang, vLLM, TokenSpeed, and KTransformers. The full GLM-5.3 weights remain a separate release.
- GPT-5.6 Sol promotion: OpenAI reduced Sol API and credit pricing for three months. Treat the promotional and list prices separately in cost models.
- Markdown for agents: Cloudflare can now serve a Markdown representation when an enabled site receives
Accept: text/markdown. It is HTTP content negotiation, not a new agent protocol and not a replacement for MCP or A2A.
That’s it for this week. Next Thursday: early community results for Qwen3.8-Flash-Next, whether GLM-5.3’s full weights arrive as announced, and the first practical reports from shared Copilot sessions.
Want this in your inbox? Subscribe to AI Dev Weekly.
Previous issue: AI Dev Weekly #22
FAQ
Is Qwen3.8-Flash-Next the same as Qwen4?
No. It is an open-weight model that previews architectural changes Qwen plans to use in Qwen4. The finished Qwen4 family has not launched through this release.
Which GPT-5.6 model should I use in Kiro?
Use Luna for high-volume routine work, Terra for the default balance of capability and cost, and Sol when difficult reasoning or architecture work justifies the premium. Validate that split on your own repository rather than assuming one model should run every task.
Does Gemini 3.5 Transcribe replace Whisper?
Not universally. Gemini reduces integration and operations work and adds managed transcription features. Local Whisper preserves offline operation and infrastructure control. Privacy, volume, latency, and required annotations decide the better fit.
Can Copilot merge code directly from Slack or Teams?
The agent can implement and validate work in a cloud sandbox and open a pull request. Repository permissions and branch protections still apply, and administrators can require an additional approval for agent-authored pull requests.