I spent the last three days reading announcements, model cards and repos, and one pattern kept showing up. The interesting releases this week are the ones that come with a check attached: a proof, a benchmark you can rerun, a test suite that passes. I have not run any of these models myself, so everything below is what the companies and authors claim, with links.
🛡️ Why can’t you use Google’s best model yet?
Google DeepMind released Gemini 4 Carbon on 28 October, just two weeks after Gemini 4 Argon. In Google’s announcement, Koray Kavukcuoglu calls it “our strongest model with state-of-the-art cybersecurity and long horizon capabilities”. All numbers here are Google’s own.
The headline figures:
- #1 on the Artificial Analysis Intelligence Index v5.0, with a score of 59.
- SEC-bench Pro 88.2% for finding vulnerabilities, ExploitGym 54.8% for exploitation, and 71.0% on Collinear’s CWE-bench for patching. Google says these were evaluated without guardrails.
- WebDev Arena up 89 Elo to #4, at 1768.
- More than 2x the success rate of Argon at generating secure code. The Chrome team reported Carbon did more than 10x better than “another lab’s large frontier model” at high-quality patches.
Google also leans hard on safety. Successful cheating on Terminal-Bench 4.0 is down to 0.0%, with cheating attempts at 0.4%. High-severity misalignment alerts in internal live coding fell almost 80%. On the Gray Swan Indirect Prompt Injection benchmark, attack success was 0.0%, against 0.7% for Argon.
Then the twist. Carbon is not generally available. It goes first to a subset of cyber defenders this week, then in phases to trusted partners through the “Fairwind Program”. Pricing is already set: an introductory $2 per million input tokens and $10 per million output, the same as Argon, with cached input 95% off. After the intro period it doubles to $4 and $20.
Why it matters for builders: cyber capability has become the thing that decides who gets a frontier model and when. A model this good at exploits helps attackers too, so Google is rationing it.
The caveat: every benchmark above comes from Google, and “as soon as possible” has no date. Budget for the $4 and $20 price, not the intro one.
✅ Can an AI prove your C code is correct?
Harmonic announced on 27 October that Aristotle, its system for research math, now does formal verification of C software and Verilog hardware. Harmonic was co-founded by Robinhood CEO Vlad Tenev. The company’s pitch line is blunt: “Your agents write more code than your team can review.”
You point Aristotle at code and state properties in plain English, or let it propose them. Each property comes back proved, refuted with a reason or counterexample, or left open, and the assumptions are listed. You commit the properties and run them in CI on every commit with aristotle ci, through the CLI and the Python SDK aristotlelib 3.0.0, released the same day. There is also a hosted MCP server.
For C, it checks functional correctness and the absence of certain classes of undefined behavior. For Verilog, it covers control-path properties like liveness and safety, plus data-path properties. Harmonic’s own case results:
- Tensordyne found bugs up to 110,000 cycles deep in its hardest block.
- In LibTomMath, the big-number library behind Tcl and Dropbear SSH, Aristotle fixed three bugs and made modular exponentiation 31 to 38% faster, with the faster code proved against the same spec.
- At the DIII-D National Fusion Facility, Mitchell Clark says one research paper plus a 254-word prompt was enough for Aristotle to reproduce a plasma control algorithm in four hours, verify the C code’s safety, and generate logically equivalent Verilog.
It is available now with self-serve, consumption-based pricing, and enterprise deployments can run on-prem or in a private cloud. Reuters notes Harmonic raised $120M last November at a $1.45B valuation.
Why it matters: if agents are writing most of your code, a machine-checked property in CI is a far better gate than a tired human reviewer. The caveat is the one from yesterday’s deep dive on Lean proofs. A proof checks the property you wrote. If the property is wrong, the proof is worthless. Someone still has to know what “correct” means.
📚 What does 3,600 pages in one pass actually need?
Baidu released Unlimited OCR 2 on 27 October. It is open source under MIT, about 3B parameters, with weights on Hugging Face, ModelScope and Baidu Cloud. The GitHub repo recommends SGLang for serving.
The new piece is Sparse Reference Sliding Window Attention, or SR-SWA. It squeezes text-visual attention over very long sequences into a fixed set of working references, for example 512 tokens, so per-step decoding latency and memory stay nearly flat no matter how long the document gets. A Page Gate narrows each query to the relevant pages, and top-scoring visual token buckets carry long-range dependencies. Xinhua compares it to a person copying a book: eyes on the current page, glancing back only when needed.
Baidu’s scores: 84.6 on MPDocBench and 77.1 on MarathonDocBench, a new hundred-page benchmark Baidu released alongside the model. Xinhua reports it ahead of Qwen 3.8-Max and Gemini-3.8-Flash, among others. On hundred-page documents, single-step decoding is 7x faster than full attention and 5.6x faster than the earlier R-SWA, per Xinhua citing the report. The demo: the full seven-book Harry Potter series, more than 3,600 pages, parsed in one pass.
Why it matters: contracts, annual reports and scanned books can now go through a small open model on your own hardware. The first Unlimited OCR repo, from 22 June, has about 29K GitHub stars.
The caveat: Baidu wrote one of the two benchmarks it leads. Test it on your own messy scans before trusting it.
Where the five stand today
| What it is | Headline number | Can you use it today? | |
|---|---|---|---|
| Gemini 4 Carbon | Google frontier model | 88.2% SEC-bench Pro | No, cyber defenders first |
| Aristotle (C, Verilog) | Formal verification service | Bugs 110,000 cycles deep | Yes, usage-based pricing |
| Unlimited OCR 2 | Open 3B document parser | 84.6 MPDocBench | Yes, MIT weights |
| Fern 0.1 | 10M-token context model | 100% single-needle at 10M | No, early access list |
| Qwenputer | CPU compiled into Qwen3.5-4B | 18M instructions, zero errors | Yes, Apache 2.0, 18.8 GB |
🧵 Is 10 million tokens of context real?
Mirendil introduced Fern 0.1 on 26 October, its first model, pretrained from scratch with a context of up to 10M tokens. That is about 23 MB of code, roughly half a million lines. The company was co-founded by ex-Anthropic researchers.
Fern 0.1 is Mirendil-Base, 0.1 to 0.5T parameters, after a short post-training round. Larger bases, up to 10T parameters, are still in pretraining. According to Startup Fortune, the company says one researcher built the base model in about 3.5 months on under $10M of compute.
The RULER results at 10M tokens, 20 examples each: single-needle 100%, multi-key 90%, multi-hop 45%. Mirendil’s own table puts its agentic coding scores behind frontier models. Mirendil also edited the post on 27 October to correct two figures.
Why it matters: if it holds up, whole codebases fit in one prompt. The caveat is that 45%. Finding a needle in 10M tokens is retrieval. Reasoning across 10M tokens is a different skill, and multi-hop is where Fern drops off. There is no API, price or weights, only an early-access list. Treat all of it as claims until outside evals arrive.
🖥️ Wait, there’s a CPU inside Qwen?
Developer Divyansh Agrawal compiled a 32-bit RISC-V CPU (RV32IM plus Zba and a bignum coprocessor) into 11 extra transformer layers placed in front of Qwen3.5-4B’s 32. No weights are learned. A compiler computes every one, and Qwen’s own weights stay unchanged. The code is Apache 2.0.
The clever bit is memory. Qwen3.5’s Gated DeltaNet state works as RAM, because the delta rule overwrites one row at a time exactly. Each row becomes a memory cell holding code, data, registers or the program counter. It runs C and MicroPython inside forward passes, passes the official riscv-tests, matches QEMU, and ran 18 million instructions free-running without error, per the author.
Qwen calls it through its normal tool-call format, but nothing outside the model executes the call. The computer layers run it and the model writes the tool response digit by digit. Agrawal reports about 3,200 trace tokens per second on an NVIDIA L4. The checkpoint is 18.8 GB, roughly double base Qwen, mostly zero padding.
Why it matters: exact arithmetic with no sandbox, and a real hint at hybrid models where some layers are compiled and some are learned. The caveat: it is a solo experiment, not Linux-class, and the numbers are the author’s.
What I’m watching next
Carbon’s wider rollout matters most, and so does whether the $4 and $20 price makes it a daily driver or a special-occasion model. For Fern, I want independent long-context evals and actual access, especially on multi-hop. For Aristotle, the question is which languages come after C and Verilog. The weekly from 19 to 25 October covers Anthropic’s side of it.
