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.
Five stories this week that actually matter. Google shipped three Gemini models in one day, including one with built-in computer use. Moonshot AI released the largest open-weight model ever. Poolside dropped a 118B model that beats DeepSeek V4 at 14x smaller. Alibaba banned Claude Code company-wide after steganographic tracking code was found. And somewhere in the noise, Gemini 4 pre-training started.
1. Poolside Laguna S 2.1: 118B that beats DeepSeek V4
On July 21, Poolside released Laguna S 2.1, their mid-tier model. This is the bigger Poolside story, not the XS 2.1. S 2.1 has 118B total parameters with 8B active, uses MoE architecture, and it beats DeepSeek V4 Pro on Terminal-Bench while being 14x smaller.
The numbers:
- 118B total / 8B active (MoE with 8 experts)
- OpenMDW-1.1 license (permissive, commercial use)
- 262K context window
- RLCEF training (same as the rest of the Laguna family)
- Runs on a single 80GB GPU with quantization
The benchmarks that matter:
- Beats DeepSeek V4 Pro on Terminal-Bench (75.2% vs 74.8%)
- Competitive with GPT-5.6 Terra on SWE-bench
- Significantly better than XS 2.1 across all benchmarks
My take: This is the model to watch. An open-weight 118B model that beats DeepSeek V4 Pro while running on a single GPU? That changes the calculus for anyone building with open models. The RLCEF training (reinforcement learning from code execution feedback) is what makes it work. Poolside is not just releasing models, they are releasing models that actually run code during training. See our Laguna S 2.1 complete guide for the full breakdown.
2. Alibaba bans Claude Code: the steganography fallout
Starting July 10, Alibaba banned all employees from using Claude Code. The reason: security researchers found steganographic markers hidden in Claude Code that flagged users in Chinese time zones. Anthropic never disclosed this tracking. Alibaba called it "high-risk software" and ordered everyone to uninstall all Anthropic tools.
What happened:
- Security researchers discovered hidden tracking code in Claude Code (since version 2.1.91, April 2026)
- The code used steganography to hide markers that detected Chinese users
- Anthropic claimed it was for "anti-resale and anti-distillation" purposes
- Alibaba banned Claude Code company-wide starting July 10
- Employees ordered to switch to internal tool Qoder
- Story hit 208 points on Hacker News
My take: This is the AI cold war going code-deep. Anthropic hid tracking code that specifically targeted Chinese users, then claimed it was about preventing unauthorized resale. Alibaba's response was predictable: if you hide surveillance code in your tools, enterprises will ban your tools. The real damage is to trust. Every enterprise security team is now asking: what else is hiding in our AI tools? This story is not going away. See our Claude Code steganography analysis for the technical details.
3. Gemini 3.6 Flash: cheaper, faster, computer use built-in
On July 21, Google released Gemini 3.6 Flash, the successor to 3.5 Flash. Same API, same SDK, just change the model name. But the improvements are real.
The numbers:
- $1.50/$7.50 per 1M tokens (output dropped from $9.00)
- 304 tok/s output speed (up from 289)
- 1M context window (same as 3.5 Flash)
- March 2026 knowledge cutoff (up from January 2025)
- 17% fewer output tokens to accomplish the same tasks
- Computer use built-in (83.0% OSWorld-Verified)
- GitHub Copilot integration available same day
The benchmarks:
| Benchmark | 3.6 Flash | 3.5 Flash | Change |
|---|---|---|---|
| DeepSWE (Datacurve) | 49% | 37% | +12 points |
| MLE Bench | 63.9% | 49.7% | +14.2 points |
| OSWorld-Verified | 83.0% | 78.4% | +4.6 points |
| GDPval-AA v2 | 1421 | 1349 | +72 Elo |
My take: The built-in computer use is the real story. Google went from "no computer use" to "83% OSWorld" in one release. That is not incremental, that is a new capability. The 17% token efficiency improvement is the kind of thing that does not sound exciting but saves real money at scale. See our Gemini 3.6 Flash complete guide for the full breakdown.
Also released: Gemini 3.5 Flash-Lite at $0.30/$2.50 (350 tok/s, fastest Gemini model) and Gemini 3.5 Flash Cyber (cybersecurity specialist, limited access).
4. Kimi K3: the largest open-weight model ever
Moonshot AI released Kimi K3 on July 16. It is a 2.8 trillion parameter open-weight model that scored #3 on the Artificial Analysis Intelligence Index, comparable to Claude Opus 4.8 and GPT-5.5.
The numbers:
- 2.8T total parameters (largest open-weight model ever)
- ~200B active parameters (estimated)
- $3/$15 per 1M tokens via API
- 1M context window
- Always-on reasoning mode
- Open weights releasing July 27, 2026
My take: The 2.8T parameter count is massive, and the benchmark scores are real. But it costs $3/$15, which is 2x more expensive than Gemini 3.6 Flash. And it runs at ~80 tok/s, which is 4x slower than Gemini. The open weights (releasing July 27) will be interesting for self-hosting, but running a 2.8T model requires serious hardware. For most developers, the API is the practical option.
5. Gemini 4 pre-training started
Buried in the 3.6 Flash announcement: Google confirmed that Gemini 4 pre-training has started. They called it their "most ambitious pre-training run yet."
My take: This is a signal, not a product. Gemini 4 is 6-12 months away. But it tells you where Google is investing. They are not conceding the frontier to Anthropic and OpenAI. They are just slower to ship.
Quick hits
- Gemini 3.5 Pro coming "soon" (delayed again). Google is a generation behind on the Pro tier.
- GitHub Copilot adds Gemini 3.6 Flash. Same day as the release. Google is getting faster at distribution.
- Poolside Laguna XS 2.1 also released July 21. 33B total, 3B active, MIT license. See our complete guide.
- Kimi K3 weights releasing July 27. If you want to self-host a 2.8T model, start planning your GPU budget now.
- Muse Spark 1.1 (July 9) still the newest Meta model. First paid Meta model at $1.25/$4.25.
What I'm watching next week
- Kimi K3 open weights (July 27). The largest open-weight model ever. How fast can the community deploy it?
- Gemini 3.5 Pro. Google says "soon." Will it close the gap with GPT-5.6 Sol?
- Alibaba's response. The Claude Code ban is just the start. Will other Chinese companies follow? Will Anthropic remove the tracking code?
- Poolside S 2.1 adoption. An open-weight model that beats DeepSeek V4 on a single GPU. If the community validates the benchmarks, this changes the open-source landscape.
AI Dev Weekly publishes every Thursday. Subscribe for the newsletter version.
Originally published at https://www.aimadetools.com
