orange tomato
Page 1/10

October's Deep Dive:
Models: Promises
and Breakthroughs

the notebook

Thirteen launches in twenty-six days. Every one of them a

sentence made of several claims. This notebook is the month

in pages: what was promised, what can be inspected, and

what changes the judgment.

Sources inspected to 28 October 2026. Desk research; no model testing.

orange tomato
orange tomato
Page 2/10

The month at a glance

1. October 2026

  • 6 Oct — Mistral Large 4 preview. Weights "by the end of the month". Still missing on the 23rd.
  • 7 Oct — Claude Haiku 5.5. The small model everyone benchmarked against next.
  • 12 Oct — Meta Muse Wonder and Wonder Full. Independently confirmed jump.
  • 13 Oct — Gemini 4 Argon. Plus a build "without cyber guardrails" for trusted defenders.
  • 17 Oct — DeepSeek V4.1-Flash-1017 weights (MIT).
  • 17–18 Oct — three "aurora" IDs appear in OpenAI's API and vanish. Never announced.
  • 19 Oct — Mercury 3: "4–8× faster at 1,000+ tokens per second".
  • 19 Oct — MiniMax M3.1 Flash: the stealth "Space Bunny" gets a name.
  • 20 Oct — DeepSeek DeepInfer, open-sourced production inference.
  • 21 Oct — Recursive Flint: a laptop coding model "largely built by AI".
  • 23 Oct — Moonshot Kimi K3.1 Pro. Capability removed first. Weights promised Nov 4.
  • 26 Oct — OpenAI GPT-6.1 Terra. "Critical in cybersecurity", per OpenAI.
orange tomato
Page 3/10

Kimi K3.1 Pro

1. What was removed

  • Moonshot says it cut cyber-offensive training data before an irreversible release.
  • Its in-house benchmark: 90.5% without the measures, 81.0% with. The "without" build never ships.
  • Reuters: "the results have not been independently verified."

2. Same benchmark, two verdicts

  • Moonshot's table: 65.7% on Terminal-Bench 4.0, "narrowly behind" Opus 5.5.
  • Artificial Analysis' own run: 67% — "the highest score of any model".
  • Both true. Different harnesses, different comparison columns.

3. One month on

  • No weights on Hugging Face. No license published. Nov 4 is a promise.
  • The cut lowered a level. It did not remove a capability.
orange tomato
Page 4/10

Mercury 3

1. The sentence

  • "The most capable diffusion-based LLM released to date."
  • "4–8× faster at 1,000+ tokens per second."

2. The picture a month later

  • OpenRouter live median: about 241 tokens per second. A quarter of the headline.
  • Artificial Analysis still has no Mercury 3 page at all.
  • The 50%-off launch price is still running. No end date posted.

3. What the harness found

  • Speed real: six times GPT-6 Luna's output rate.
  • But 7.5–8.6% of code calls reasoned silently and returned nothing.
  • Cheap per token is real. Cheap per correct answer is a different question.
orange tomato
Page 5/10

MiniMax M3.1 Flash

1. The five layers

Identity — the stealth model gets a name.
Usage — traffic, not quality.
Vendor benchmarks — two names, two instruments.
Price — 25% or 50% of M3, not one third.
Weights & license — downloadable, under conditions.

2. One month on

  • The card says 428B parameters. The same page's sidebar says 220.49B.
  • The free window ended 22 Oct. Same weights: 31 to 166 tok/s by host.
  • 62.8 and 72.5 were different benchmarks with near-identical names.
orange tomato
Page 6/10

The rest of October

1. Kept brief

  • DeepInfer (20 Oct): the production stack, open-sourced. KV cache, routing and load balancing are not in the box.
  • Flint (21 Oct): TechCrunch measured ~131 tok/s on a laptop. The "AI builds AI" claim has no auditable record.
  • aurora: three unannounced API IDs and some cat portraits. Identity remains a guess.
  • Mistral's weights: promised "by the end of the month". Unpublished by the 23rd.
  • Argon's second build "without cyber guardrails to a select set of trusted defenders".
  • GPT-6.1 Terra: OpenAI's own card calls it "Critical in cybersecurity".

2. The shape of it

  • Capability and safety statements now travel inside the launch sentence.
  • They are, almost always, the vendor's own assessments.
orange tomato
Page 7/10

The pattern

1. One sentence, several claims

The launch
sentence
→
1. Who ran the test?
2. Which instrument, which version?
3. What does the price leave out?
4. What does the license say?
5. What does live traffic show?

2. The instrument moves too

  • The same model's index page read 43 at launch week, 29 five days later. The index changed version.
  • Numbers without a version and a date are not reproducible.
orange tomato
Page 8/10

What you can inspect

1. The checklist

  • Who ran the test — vendor, evaluator, platform, or a community harness?
  • Which instrument, at which version? Pin the benchmark name and the harness.
  • What does the price exclude? Retries, thinking tokens, tools, cache, the promo's end date.
  • What does the license actually say? "Open weights" describes files, not permissions.
  • What does live traffic show today? Platform medians are the user's experience.
  • Promised or shipped? Weights "by Nov 4" and weights in a repository are different facts.
  • What is the sentence not saying? Every launch sentence was true of a narrower claim.

Editorial guidance from the month's evidence, not a proven standard.

orange tomato
Page 9/10

What changes
the judgment

1. The evidence that does not exist yet

  • Nov 4: Moonshot's weights, the license, the first independent tests.
  • An Artificial Analysis Mercury 3 page.
  • A reproducible cost-per-correct-answer measurement.
  • A parameter census of MiniMax's files: 428B or 220.49B?
  • Mistral's promised weights.

2. Until then

  • The launches are real. Several gains are independently confirmed.
  • Every headline sentence is bigger than the evidence behind it.
  • Only some of its parts come with receipts.
orange tomato
Page 10/10

Sources & method

1. What this notebook is

  • Desk research: published documents, inspected to 28 October 2026.
  • No model was run, queried, or tested. Every figure is attributed to whoever measured it.

2. Where the claims live

  • Launch blogs: Moonshot, Inception, MiniMax, Mistral, Meta, Google, OpenAI.
  • Independent checks: Artificial Analysis, OpenRouter live statistics, the community Mercury 3 harness, Plotly's data-analytics exam.
  • Files: Hugging Face model cards, licenses, config.json.
  • Full links and the claim ledger live in the Deep Dive and its run folder.

3. Limits

  • In-house benchmarks cannot be audited. Promises are not shipments.
  • Moving figures are dated. Where the record stops, the notebook says so.

Companion to: October's Deep Dive: Models: Promises and Breakthroughs.

orange tomato