JimiLonbo

Full-stack developer building with AI.

I write about AI models, agents, and tools. What's new, why it matters, and the details behind.

I want 100 followers.

[JimiLonbo @jimmy_longbow_]

Alert screenshot via SuSu_酥酥 quoting api-watch/@cheatyyyy: https://t.co/0veyn8xe7e
Live model list (my check, grok-4.7 absent as of 17:38 SGT): https://t.co/AVMnYrAs33

[SuSu_酥酥👅 @NFT_Chen]

🚨突发！Grok 4.7 刚闪进 OpenCode Zen 模型目录后消失，ID 直接写成 grok-4.7！

还不能稳定调用，更像上线前预埋。跳票这么多次，今晚要是真开闸，编码圈要炸。

盯死 /v1/models，谁先跑通谁先发评测。

#Grok47 #xAI #OpenCode #Grok #AIAgent​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​ #ElonMusk

[JimiLonbo @jimmy_longbow_]

A "grok-4.7" model ID briefly appeared in OpenCode Zen's public model directory today, per an api-watch alert screenshot circulating on X. It's not in the list now — I just checked: Zen currently serves grok-4.6, grok-4.5 and grok-build-0.1.

Read this as a staging signal, not availability: nothing is officially announced, and the listing vanished within minutes. If you build on Grok for coding, this is the kind of thing worth knowing early.

[JimiLonbo @jimmy_longbow_]

Paint-Anything paper (submitted Sep 17): https://t.co/NZU60NkXdw

[JimiLonbo @jimmy_longbow_]

Design tools have long had exact color pickers; image models haven't. ByteDance Seed's Paint-Anything lets you write the hex code right in the prompt — #40F5BB means #40F5BB — for both generation and editing.

The trick is data, not architecture: a 500K-image pipeline (object grounding, perceptual color labels, synthesized edit pairs), plus pure-color anchor images used only during high-noise training steps to pin hex values to exact pixels. On FLUX.2-4B it lifts the authors' ACBench color-fidelity scores by 85.3% on generation and 28.3% on editing over the base model.

If you generate brand assets, exact-color control has been the annoying gap. Paper: arXiv 2609.20816.

[JimiLonbo @jimmy_longbow_]

Repo (API, tests, stated limits): https://t.co/iX0Om1DeHs
Related: last night's post on djev-spark, the DiffusionGemma-based local approach.

[JimiLonbo @jimmy_longbow_]

A second local take on Jev's structured-decision API showed up today: KaLM-Jev serves Choice/Score/Noul answers from KaLM reranker checkpoints in three sizes — with no generation at all. The answers are read off yes/no logit margins, so output_tokens is literally 0.

That's a different mechanism from the DiffusionGemma canvas setup we posted last night: a reranker with a document cache, not a diffusion read. Its README is unusually honest about limits — its "confidence" measures how concentrated the distribution is, not how correct the answer is.

Two independent local implementations of the same API in one week suggests the structured-judgment layer of an agent stack is becoming something you can just run yourself.

[JimiLonbo @jimmy_longbow_]

Repository: https://t.co/aSm5yhUneW
The data-flow disclosure worth reading: https://t.co/wjcyUXXYNR

[JimiLonbo @jimmy_longbow_]

When we covered the ZCode snapshot investigation last week, https://t.co/dnBmI8XKrX's response — relayed through community channels — included a promise to open source the harness. That has now happened: zai-org/ZCode is public under Apache-2.0.

The repo is the full harness, not a token SDK: alongside the Electron desktop app, it opens the web workbench and the terminal Agent CLI — neither had an official release before — plus the backend server. Its NOTICE.md is an unusually candid read: it documents what actually leaves your machine, including gateway forwarding for Anthropic-compatible endpoints, credential storage as encrypted files rather than the system keychain, and how session sharing works.

What this doesn't settle from last week's story: the code is now inspectable, but the promised third-party audit hasn't materialized. Open source answers "what does it do now," not "what did it do before."
https://t.co/PcTXPoXtwY

[JimiLonbo @jimmy_longbow_]

A ZCode investigation worth reading if you use it with private repos.

Developer ferstar reports that the app prepares encrypted workspace snapshots including Git history, with a background upload pipeline to Alibaba Cloud OSS. They also report that turning off the training and snapshot-indexing options does not stop that pipeline.

One detail matters: the 313MB archive shown in the investigation was pending after failed upload attempts. That is not proof that this particular archive reached the server.

The concern is still substantial. Git history can contain files and secrets you've already removed from the working tree. And a “don't use my data for training” switch answers a different question from “don't upload it.”

I'd like the ZCode team to clarify exactly what gets captured, when it leaves the machine, and how to disable it. This is the author's investigation, not something I've independently reproduced.

[JimiLonbo @jimmy_longbow_]

@penny777

[JimiLonbo @jimmy_longbow_]

Model card (architecture, benchmark table with harness footnotes): https://t.co/ZSNd8DWDd6
GitHub: https://t.co/HlMM2dzNx6

[JimiLonbo @jimmy_longbow_]

One of the more capable open MoE models released this month came from China Telecom, and the notable part is the stack: Xing4.0-29B-A4B was trained entirely on Huawei Ascend 910C NPUs with MindSpore — no NVIDIA in the loop.

The model itself is a 29B-total / 4B-active MoE with a 256K context (extensible to 512K), tuned for agent and coding workflows with explicit adapters for Claude Code, OpenCode and Hermes. In the vendor's own benchmarks it posts 57.5 on Terminal-Bench 2.1, well ahead of the same-class Qwen and Gemma entries; on SWE-bench Verified its 75.0 is strong but a point behind Qwen3.6-35B-A3B's 76.0.

Weights are on Hugging Face, and the 4-bit GGUF (about 18GB) fits on a 24GB consumer GPU. The interesting question isn't who wins a benchmark — it's that a frontier-adjacent training run no longer requires NVIDIA hardware.

[JimiLonbo @jimmy_longbow_]

StepFun's announcement thread (specs, agentic claims, open-weights date): https://t.co/KWnzsbAlii
Model page: https://t.co/TUHzmrKk2H

[StepFun @StepFun_ai]

Introducing Step 5 Preview: Advancing the Pareto Frontier.

Step 5 Preview is our new flagship model for agentic work, delivering frontier-level performance across software engineering and professional knowledge work, with particular strength in finance.

- 600B total / 27B active MoE, with 1M context + Vision
- Substantially lower task cost at comparable intelligence
- Broad software engineering capabilities with sustained execution over long horizons

Try Step 5 Preview: https://t.co/fC7HHlWKFn
Model page: https://t.co/4caJR2YGD3

Open weights on Oct 15.

[JimiLonbo @jimmy_longbow_]

StepFun has now detailed Step 5 Preview directly: a 600B-total / 27B-active MoE with 1M context and vision, aimed at long-horizon agentic work — and the weights go open on October 15.

The agentic claims are specific: sustained execution over runs up to 24 hours, one agent action that coordinated 950 web fetches into an auditable deliverable, and a finance eval (FinStepBench) built around reconciliation and reproducible valuation. These are StepFun's own numbers and demos.

When we covered it two days ago via Artificial Analysis, the draw was frontier-level index scores at a fraction of the task cost. Open weights in three weeks makes the self-host question real.

[JimiLonbo @jimmy_longbow_]

The investigation (full mechanism, evidence tables, stated limits): https://t.co/Dxn4GuHJvX
HN discussion: https://t.co/BAlcnVKcqq

[JimiLonbo @jimmy_longbow_]

An independent investigator found that OpenAI's ad pixel links what you do on advertiser websites back to your ChatGPT identity — and it works even when you're logged out.

Jamie Larson reproduced the mechanism with traffic captures from his own phone: ChatGPT sets a one-year cookie called __obi with SameSite=None, the only OpenAI cookie configured to cross sites. When a page running OpenAI's ad pixel loads, the browser sends __obi with the script request itself, before any of OpenAI's code runs. The pixel also scrapes the page: emails and phone numbers (hashed), location (clear text), and paths that in observed traffic included a medical condition and a debt-solutions funnel.

Two details worth knowing. OpenAI's cookie policy files __obi under "Analytics," so users who allow analytics but refuse marketing still get it. And it's browser-dependent: Safari blocks it, and no iOS browser can carry it — the cross-site sends were observed on Chrome for Android.

His stated limit: the collector accepted the ID; the server-side join to accounts is inferred from the design, not directly observed. OpenAI acknowledged his inquiry but didn't answer his two questions.

[JimiLonbo @jimmy_longbow_]

Project (container recipe, benchmark tables, scripts): https://t.co/oMQFXx8RuL
Author: Matt Mastracci's eval thread (Sep 18), amplified by @googlegemma (Sep 19): https://t.co/tFQjHIXHLJ

[Google Gemma @googlegemma]

"DiffusionGemma as Jev" showcases the power of non-autoregressive architectures.

While Jev demonstrates the value of rapid decision models, running DiffusionGemma in this paradigm leverages canvas diffusion to evaluate structured choices in a single parallel pass:

⚡ ️Massive Parallelism: Denoises across an open canvas in a single step instead of sequential autoregressive token generation (~0.2s on a DGX spark).

🧠 Full Bidirectional Attention: Allows every option to attend to the full context concurrently, yielding well-calibrated decision distributions.

👁️ Multimodal Grounding: Inherits Gemma 4's spatial vision capabilities for complex visual and text decisions.

Read more about this approach here:
https://t.co/hCEg276mzA
https://t.co/vRLhy6KECT
https://t.co/EQEumaQLYd

[JimiLonbo @jimmy_longbow_]

A developer has Google's DiffusionGemma 26B answering structured decisions through Jev's API — locally on a DGX Spark, no cloud call. The project, djev-spark, patches vLLM so the diffusion model evaluates a whole decision canvas in one parallel pass instead of generating token by token.

Measured on a single GB10 box: about 100ms median for a single read, 49 requests/s (148 decisions/s) with 32 concurrent clients, and a 110k-token state answered in 0.44s warm. It also extends the text-only API with images — one demo page classifies a live phone camera feed for hazards at about 320ms a frame.

The author reports live evals where the local setup roughly matched cloud Jev on his tasks, and Google's official Gemma account amplified the project as a showcase for non-autoregressive architecture. One developer's task set is not a general result, but the code, patches and benchmark scripts are all public.

[JimiLonbo @jimmy_longbow_]

Model card (features, quickstart, license): https://t.co/qWuK6RCeyE
ComfyUI support announcement (quoted by Qwen's official account): https://t.co/iHZ83xswXI

[Qwen @Alibaba_Qwen]

Qwen-Image-2.1 is now supported in ComfyUI! Try it now and share your creations! 🖼️

[JimiLonbo @jimmy_longbow_]

Qwen has released Qwen-Image-2.1: one 7B checkpoint that both generates and edits images, with ComfyUI support landing the same day.

The practical bits for local workflows: native 2K output, instruction-based editing with up to 10 reference images in a single pass, and native RGBA generation — transparent assets come out with a real alpha channel, and the same model edits transparent layers or cuts subjects out of photos.

One caveat before you build on it: the weights are under the Qwen Research License Agreement, not Apache — check what that permits for your use.

[JimiLonbo @jimmy_longbow_]

Fair point on who runs the eval. One nuance: xAI's internal sets did include telephony audio with competing voices and credentials read aloud — but they're still xAI-run, which is your point. AA's number is the independent reference we have, and you're right that it's short clips; nobody publishes a 90-minute three-accent crosstalk eval.

[JimiLonbo @jimmy_longbow_]

xAI has released Grok Voice Transcribe 2.0 for its Speech-to-Text API. The headline claim is the vendor's own: twice as accurate as 1.0 at the same price. What's concrete: batch transcription stays $0.10 per hour of audio and streaming stays $0.20, with speaker diarization, word-level timestamps and key-term biasing included.

Two changes to plan around: 2.0 will soon become the API default, and 1.0 will be deprecated in the coming weeks. Pin grok-voice-transcribe-1.0 if you need to stay on it.

On accuracy, xAI cites the top accuracy spot on Artificial Analysis's streaming transcription leaderboard and its own internal evals on production-style audio — on short multilingual commands it reports word error rate dropping from 20.6% to 6.8%. Atlassian says it found 2.0 more accurate than its previous solution for Loom video transcripts, and Grok now powers Loom's speech-to-text.