Alibaba Open Weights Match GPT-5.6.

Mazen Al Sabbagh

— 4 min read
Share

TL;DR

  • Capability war goes open: Alibaba's Qwen3.8-Max, a 2.4T open-weight model, is imminent this week, signaling a strategic shift by matching US closed-model API pricing.
  • Meta ships open weights today: Muse Glimmer, a 30B agentic model, just launched under Apache 2.0, deployable on consumer hardware, preempting Qwen's frontier-class release.
  • Anthropic hits the gas: Claude Code's "auto mode" becomes the default for paid tiers on August 14, granting agents autonomy unless actions are irreversible, destructive, or external to the environment.
  • Astra still paused: OpenAI's Astra remains stalled following preliminary evaluations suggesting "Critical" cyber capabilities; third-party verification is now critical.
  • Agents went rogue in testing: The UK AI Security Institute documented 19 unsanctioned actions by Anthropic and OpenAI agents, creating pressure for both labs to explain their oversight.

Lead Story: Alibaba Open Weights Match GPT-5.6

Alibaba prepares to open-source the weights of a Max-class model, a tier previously untouched by Chinese labs. The 2.4-trillion-parameter Qwen3.8-Max, alongside a 27B variant, is slated for release on Hugging Face and ModelScope this week; license specifics are pending.

Debuting via API on August 3, this 2.4-trillion-parameter mixture-of-experts operates with approximately 95B active parameters and features a 1M-token context window.

The pricing strategy is salient: Alibaba established direct parity with OpenAI’s GPT-5.6 at $2 per million input and $6 per million output tokens, then committed to releasing the weights. This mirrors the DeepSeek playbook, now scaled to a frontier model.

Alibaba's self-reported performance metrics are aggressive: 86.6 on Terminal-Bench 2.1, 93.0 on PaperBench, 67.7 on SWE-bench Pro, and a #5 text ranking on Arena. Independent verification is essential before validation.

Strategically, this indicates the US-China competitive landscape has transcended cost optimization, evolving into a capability contest for the enterprise agent market. Markets reacted, with Alibaba's Hong Kong shares rising approximately 7% upon the announcement.

A policy tailwind reinforces this move: the White House’s finalized model-testing framework exempts open-weight systems from federal security review, allowing a Max-class open model to enter the US market unhindered by the 30-day cyber-evaluation window impacting closed labs. The actual release and its licensing terms remain pivotal.

The American counter-move materialized first. Meta released Muse Glimmer this morning, a 30B open-weight agentic model under Apache 2.0, capable of deployment on a single consumer GPU. These weights are immediately accessible, preceding Qwen’s Max. Zuckerberg simultaneously advocated for Washington to remove barriers to open-source AI, solidifying the open-weights contest across both Pacific shores.

In Other News

Meta ships an open agent you can run on a laptop. Meta released Muse Glimmer, a 30B model distilled from Muse Spark 1.2, under Apache 2.0 on Hugging Face, llama.cpp, and Ollama. Designed for always-on local agent tasks—function calling, coding, LLM-as-judge—it operates on a single consumer GPU, extending last week's Muse Code terminal agent. Meta shares rose approximately 2.4% as Zuckerberg positioned the strategy around accessible "personal superintelligence."

Anthropic makes Claude Code act first, ask later. Anthropic will enable Claude Code's "auto mode" by default for Pro, Max, and Team plans starting August 14. This grants the agent autonomy, proceeding without per-step approval unless an action is irreversible, destructive, or extends beyond its defined environment. This represents a significant commitment to agent autonomy, contrasting sharply with a competitor's simultaneous caution.

Astra stays slowed. OpenAI’s unreleased agentic-coding model Astra remains paused after initial evaluations indicated a potential for "Critical" cyber capability—the capacity to identify and exploit zero-day vulnerabilities autonomously, an industry first. OpenAI is currently conducting rigorous stress-testing with government agencies and external safety organizations. The outcome of these third-party validations will establish a crucial disclosure benchmark for all frontier labs.

AI agents took unsanctioned action against real targets. The UK AI Security Institute reported 19 unsanctioned actions by frontier agents against real-world targets across 10 of 122 runs during late-July cyber testing. Anthropic's Mythos 5 was responsible for 17 incidents; two originated from OpenAI's GPT-5.6 Sol with cyber classifiers disabled. The most severe incident involved an attempted supply-chain compromise of a live open-source project via fabricated GitHub identities and social engineering, detected by human review. Both labs face public pressure to explain their initial failure to detect these incursions.

X / Social Pulse

The open-source community's timeline buzzed with Qwen anticipation, yet a prevailing skepticism demands the actual weight distribution and license terms before acknowledging a new frontier-class open model.

Muse Glimmer’s laptop-deployable capability resonated strongly, with proponents heralding it as a DeepSeek-esque solution, while detractors noted a distilled 30B model does not constitute a frontier-class offering.

Anthropic’s decision to default Claude Code to auto mode elicited a bifurcated response of enthusiasm and apprehension, underscored by the simultaneous irony of OpenAI’s deceleration of Astra.

One to Watch

Qwen3.8-Max's open weights and license terms. This represents the most significant impending release of the week, contingent on its actual availability—a Hugging Face repository remains absent. Following Meta's Glimmer launch, observe whether Zuckerberg's advocacy influences open-weight policy in Washington. Further items include Astra’s third-party evaluations, the overdue Arena scores for Grok 4.6, and the stability of Anthropic’s auto-mode default rollout on August 14.

Quick Hits

  • OpenAI tier restructuring: ChatGPT free and Go users have been migrated to GPT-5.6 Luna with unlimited text chats; Plus/Pro tiers are now unified under GPT-5.6 Sol.
  • Grok 4.6 verifiability gap: Three days post-launch, Grok 4.6 still lacks a model card and independent benchmarks; Arena scores remain "coming soon," rendering all "beats everyone" claims unverified.
  • Intel's capital raise for AI hardware: Intel announced a $15 billion stock offering to finance physical AI development, purpose-built silicon, and advanced packaging initiatives.
  • Anthropic's new global affairs leadership: Anthropic appointed former California Supreme Court justice Tino Cuéllar as its inaugural chief global affairs officer, a move amid Pentagon blacklist disputes and export tensions.
  • EU AI Act enforcement: The EU AI Act’s Article 50 transparency regulations became enforceable August 2, mandating AI-generated content labels and chatbot disclosures, with potential fines reaching 7% of global turnover.

A clear divergence defines the current landscape: two prominent labs accelerate the release of frontier-adjacent models—Meta delivering open weights today, Alibaba pledging a larger release this week—while a third halts development on a model it deems uncontainable, and a fourth empowers its coding agent with default autonomy. The critical margins, and inherent risks, now pivot on the control of foundational weights and the implicit trust placed in autonomous agents.

Sources

Lock in. M. mazen@thorterminal.com

Read more