October 12 was a distribution day. A model got small enough for a phone. A harness shipped a major version. And a subscription bundle put Grok, Cursor, and X Premium on one bill.

No single item is a frontier breakthrough. Together they show where the fight moved: who owns the surface where you use the model.

MiniCPM-V 4.7 put multimodal AI on a phone

@OpenBMB released MiniCPM-V 4.7 in two sizes: a 35B-A3B model and a 1B model.

The claims are aggressive:

  • 17+ tokens per second on an iPad Pro M4 with 16GB.
  • 13+ tokens per second on an iPhone 17 Pro Max.
  • A 5x speedup from Lookahead Expert Scheduling (LES), with byte-identical outputs.
  • Competitive results across 34 benchmarks.
  • 90.4 on OCRBench at 16x visual compression, 5.2 points ahead of Qwen3.6 35B-A3B.

The post also says the model costs "only 3GB memory." I have not verified that line, and I would not repeat it as fact. A 35B-A3B model needs memory for weights, KV cache, and runtime. The number may refer to a specific compressed path or an incremental buffer. Ask before you size a device around it.

The LES idea is the part I like. It predicts and preloads expert weights to avoid inference stalls. That is a systems fix, not a model fix. It also explains why the speedup can keep outputs identical: the math does not change, the scheduling does.

On-device multimodal AI has been a demo category for two years. A 35B model at interactive speed on a tablet is a different product. It means OCR, document QA, and visual assistance can run without a round trip. That is a privacy and latency win in one move.

OpenCode 2.0 and the $80 experiment

@opencode shipped OpenCode 2.0 with a short post and a link to the notes at anoma.ly/notes/opencode-2-0. The post got 8,290 likes and 799,479 views.

That day, @opencode ran "Operation Cheepseek Phase 3": $80 of usage for DeepSeek V4.1 Flash. Another 7,719 likes.

I like this more than a launch keynote. A harness that publishes a fixed budget and a model name gives you a reproducible experiment. You can ask the useful question: what can $80 of this model do inside this tool?

That is also a pricing signal. If a coding harness can hand out $80 of inference as a promotion, the marginal cost of tokens is not the constraint anymore. The constraint is attention.

The X Subscription bundles the tools

@SpaceXSI announced The X Subscription, which unifies plans across Grok, Cursor, and X Premium. The post got 17,278 likes and 11.6 million views.

That is a big number for a billing change. It tells you how many people already pay for two or three of those products and hate the separate bills.

The details that matter are not in the post: price, limits, and what happens when Cursor's usage outgrows the bundle. Read the terms before you cancel anything. A bundle is a discount until the cap hits.

The strategic read is simple. If the subscription owns the chat, the editor, and the social feed, then the model inside matters less. The bundle becomes the product. That is the same move the cable companies made, and it worked for a long time.

QwenGyre made long-horizon training cheaper

@Alibaba_Qwen introduced QwenGyre for extreme-long-horizon agent training.

Two designs do the work. An elastic scheduler reallocates GPUs between rollout and training as demand shifts, without interrupting live executions. It claims up to 1.85x speedup over Colocate and 1.78x over Async. A trajectory processor turns branching executions into prefix-shared trees with bounded, deduplicated samples. Partial progress after a timeout still gets scored.

At Qwen3.8-2.4T-A95B with about 700K tokens per rollout, QwenGyre lifted NL2RepoBench from 52.5% to 58.5% in 48 steps. It cut end-to-end training time by about 44% versus Async.

This is infrastructure work, and it is the kind that decides who can afford the next agent model. Long-horizon tasks generate messy, branching traces. If your trainer cannot use the branches, you pay for them twice.

OpenClaude grew up

@twigpine shipped OpenClaude 0.32.0 from its new home. The numbers: 33.7K stars, 392K downloads, 158 contributors, six months.

npm install -g @twigpine/openclaude is the whole onboarding. That is how you win developers. Not a waitlist.

A 2B decider that beats its teacher

@SeraAndroid posted v2 results for a Gemma 4 E2B decision model:

metricv1v2
accuracy0.5130.804
KL from gold0.3670.182
Brier0.1970.082
p50 latency0.17 s0.61 s
ECE0.0410.061

The teacher model scores 0.735, so the student now beats it. A dedicated fine-tuned decider lands at 0.792.

The honesty is the best part. Accuracy went up, latency went up 3.6x, and calibration slipped. That is a real trade-off, stated plainly. If your agent makes a routing decision on every step, 0.61 seconds per decision is a budget item, not a footnote.

The adoption gap

@1337hero made the argument I keep coming back to. Inside the tech bubble, adoption looks instant. Outside it, companies are slow, and humans make giant messes. He works with Fortune 5000 clients whose IT teams are afraid to run npm install.

His conclusion: there will still be software engineering jobs in 2028, but you do not want to be the last dinosaur.

I would add a corollary. The tools that win the enterprise will be the ones that survive a security review, not the ones with the best demo. That is boring, and it is true.

What I would do

Test MiniCPM-V 4.7 on one on-device task with a real document. Measure memory, not just tokens per second. Ignore the 3GB line until someone publishes the recipe.

Install OpenCode 2.0 and run the $80 experiment on a task you already know. Compare it to your current tool on cost per completed task, not cost per token.

Read the X Subscription terms before you switch billing. Bundles are good until the cap.

And if you fine-tune a small decider, publish the latency and calibration numbers too. Accuracy alone is a misleading score.

Sources

#minicpm-v 4.7#opencode 2.0#x subscription#on-device multimodal ai#qwengyre#openclaude
Raja CRN

Raja CRN

Tech enthusiast and developer.