Icosa claims Zeno can offload its bundled 4-bit Qwen3.6-35B-A3B to run fully locally on 16GB Macs with usable agentic file-work performance, which would make a 35B-class local work agent viable on base-memory Apple hardware.

state: expiredheat: lowuncertainty: highknownscott: mediumlocal-inference coding-agentsIcosa

What is this?

Icosa announced Zeno, described in the supplied evidence titles as a free local Mac AI work tool bundling a 4-bit Qwen3.6-35B-A3B model and claimed to run fully locally on a 16GB Mac. The supplied web snippets do not independently document Zeno or explain Icosa’s offloading method, while hardware guidance for comparable 35B-A3B 4-bit models reports roughly 18–19.5GB model memory, a practical floor of about 32GB, and substantially more memory for useful agentic context. The core 16GB claim is therefore unverified and conflicts with the available memory estimates, despite a separate article arguing that the model’s 3B active parameters make local agentic use comparatively viable.

Why it matters to Scott

The radar already tracks this exact development in `radar:zeno-qwen35-16gb-mac`, including the need for independent testing of practical agentic performance. The claim bears directly on Scott’s hardware-aware local-inference work and the local execution path in his Ask terminal agent, but the supplied evidence adds no verification that would yet change his build choices.

dev:concept.hardware-aware-local-inferencedev:project.askradar:zeno-qwen35-16gb-macradar:concept.local-inferenceradar:concept.apple-silicon-inferenceradar:concept.coding-agents

queries asked of Scott's wikis

local-agent memory budgeting and context limits
Apple unified-memory inference constraints
MoE active parameters versus total model footprint
model offloading for low-memory local inference
minimum viable hardware for local coding agents
local-first agentic file-work architecture

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (2) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	We built a local AI work tool that runs Qwen3.6-35B-A3B on a 16GB Mac (update) LocalLLaMA	close_Meal6005	0	9
🟧 echo.other ⭐	The earlier post is the same author's original announcement: Icosa released Zeno, a free local Mac AI work tool using a 4-bit 35B Qwen model	close_Meal6005 (self-described Icosa intern)	—	—

Interpretation history

2026-08-28T19:36:22Z
No independent testing or implementation evidence appeared within the observation horizon, leaving the consequential 16GB agentic-performance claim unverified. The low-traction announcement has faded without establishing an active developing episode.
2026-08-26T19:31:53Z
Refreshed comments reinforce trust and real-workload concerns but add no independent measurements or hands-on results. The app remains established, while its consequential claim of usable 35B-class agentic work on a 16GB Mac is still unverified.
2026-08-26T17:42:14Z
Refreshed discussion adds skepticism and interest in real-workload testing, but no independent benchmark or implementation evidence. Zeno’s existence is established while the consequential 16GB agentic-performance claim remains unverified.
2026-08-26T17:31:27Z
grounded: known/medium — The radar already tracks this exact development in `radar:zeno-qwen35-16gb-mac`, including the need for independent testing of practical agentic performance. Th
2026-08-26T17:28:09Z
origin walked (codex/luna, conf 0.96): anchor reddit.post.1vz2d9f -> echo.other.dd8804fa6c by close_Meal6005 (self-described Icosa intern)
2026-08-26T17:26:59Z
case created — This is a concrete local-agent release making a notable constrained-hardware claim distinct from existing Qwen3.6 inference episodes.

Decision trace

08-29 05:36expireNo independent testing or implementation evidence appeared within the observation horizon, leaving the consequential 16GB agentic-performance claim unverified. The low-traction announcement has faded
08-29 05:36alert_silentThe only delta is elapsed staleness, not new evidence; Scott can revisit if concrete throughput, memory, context-length, or task-completion results emerge.
08-29 05:36alert_routeThe only delta is elapsed staleness, not new evidence; Scott can revisit if concrete throughput, memory, context-length, or task-completion results emerge.
08-27 05:31repriceRefreshed comments reinforce trust and real-workload concerns but add no independent measurements or hands-on results. The app remains established, while its consequential claim of usable 35B-class ag
08-27 05:31alert_silentThe delta is repetitive discussion rather than evidence about memory pressure, context capacity, throughput, or task completion; it can wait for independent testing or concrete benchmarks.
08-27 05:31alert_routeThe delta is repetitive discussion rather than evidence about memory pressure, context capacity, throughput, or task completion; it can wait for independent testing or concrete benchmarks.
08-27 04:21sensor_dirtycomment_update
08-27 03:42repriceRefreshed discussion adds skepticism and interest in real-workload testing, but no independent benchmark or implementation evidence. Zeno’s existence is established while the consequential 16GB agenti
08-27 03:42alert_silentThe new comments are low-engagement discussion rather than evidence about throughput, context limits, memory pressure, or task completion, so they do not change Scott’s decisions and can wait for conc
08-27 03:42alert_routeThe new comments are low-engagement discussion rather than evidence about throughput, context limits, memory pressure, or task completion, so they do not change Scott’s decisions and can wait for conc
08-27 03:40alert_silentThis repeats the already tracked first-party release and 16GB-Mac offloading claim. The only apparent addition is unspecified vendor-run comparison testing in an inaccessible image; it provides no ind
08-27 03:40alert_routeThis repeats the already tracked first-party release and 16GB-Mac offloading claim. The only apparent addition is unspecified vendor-run comparison testing in an inaccessible image; it provides no ind
08-27 03:31groundThe radar already tracks this exact development in `radar:zeno-qwen35-16gb-mac`, including the need for independent testing of practical agentic performance. The claim bears directly on Scott’s hardwa
08-27 03:28promote_anchororigin walk conf 0.96
08-27 03:27createThis is a concrete local-agent release making a notable constrained-hardware claim distinct from existing Qwen3.6 inference episodes.