AI Engineering Digest

Ollama 0.34.1 ships MLX create GA.

Wednesday, September 16, 2026

Issue 18 · sourced briefing

Sources: Ollama · GC RN · AWS ML · Claude RN · llama.cpp · PEFT

Breaking Ollama v0.34.1 is GA — MLX safetensors create leaves experimental; Issue 17’s RC closes

Lead Story

Ollama 0.34.1 GA — MLX create leaves experimental

Stable v0.34.1 (~4:10 PM ET Sep 15) promotes MLX safetensors ollama create out of experimental, tightens Apple Silicon MLX memory handling, and speeds large-library /api/tags — closing the RC-only note from Issue 17.

GitHub marks v0.34.1 as Latest (atom updated 2026-09-15T20:10:42Z). Release notes: MLX safetensors ollama create is no longer experimental; GGUF creation now requires llama.cpp tooling for safetensor conversion and quantization. Also: improved MLX memory on Apple Silicon; runaway repeat detection raised to 100 tokens (fewer OCR false positives); /api/tags faster on large libraries (3.1 s → 294 ms cold in testing — vendor test); typical_p deprecated for new model creates (existing GGUF retains support); MLX and llama.cpp updates.

Issue 17 only carried v0.34.1-rc2 as a prerelease. Do not confuse with v0.34.2-rc0 (same evening) — still an RC with “llama.cpp updates” only. Stable baseline before this was v0.34.0 (Issue 14).

Why it matters: Mac operators can treat MLX create + memory behavior as GA, not an RC experiment, and upgrade without waiting on another cut.

Sources: Ollama v0.34.1 · Ollama releases

Today’s Stories

Enterprise · ML practice

RL fine-tuning lands in the Agent Platform console

Google Cloud RN Sep 15: create, monitor, and test Gemini reinforcement learning fine-tuning jobs in the console (Preview) — reward functions, live metrics, Agent Studio checkpoint tests. Pre-GA: not for production/proprietary data. vendor

Source: GC RN · RLFT docs

Enterprise · SageMaker

SageMaker prefers a list of GPUs, not one type

AWS blog Sep 15: Training and Processing jobs take an ordered list of up to five instance types and launch on the first with capacity — with Flexible Training Plan priority and on-demand fallback. Cuts manual retry scripts.

Source: AWS ML blog

Enterprise · security

AgentCore Consent portal binds OAuth for gateways

Bedrock AgentCore Identity (blog Sep 14): managed Consent portal + session binding for Gateway — corporate IdP sign-in, per-provider GitHub/Slack grants, tokens in the Identity vault. Aimed at IDE/MCP clients. vendor

Source: AWS ML blog

Frontier · Claude

Messages API can compact on demand (beta)

Claude Platform RN Sep 14: beta header compact-2026-09-04 — send compaction, get a signed summary block, reuse it instead of full history. You choose when; can keep recent turns verbatim (thinking may stay valid on preserved-thinking models).

Source: Claude Platform RN

Local · Metal

llama.cpp b10994 stops Metal MoE NaN prefills

Nightly fixes mul_mm_id f16 overflow → NaN when activations exceed f16 range (e.g. Mistral Small 4 MoE). Author benches M2 Max only — not a Studio claim. Dense mul_mm noted as same risk, not fixed here. contributor

Source: b10994 · #26223

ML practice · PEFT

PEFT 0.21.0 adds KaSA, Super-Tuning, ShadowPEFT

Hugging Face PEFT v0.21.0 (~11:31 AM ET Sep 15): Riemannian-preconditioned LoRA optimizer, KaSA, Super-Tuning, ShadowPEFT, MoE speedup, LoRA combine SVD double-scale fix. MetaMathQA memory/time figures are vendor.

Source: v0.21.0

Signal Tables

Model & runtime releases

Model / package	Org	Notes
Ollama 0.34.1 Sep 15 GA	Ollama	Open MLX create GA · memory
PEFT 0.21.0 Sep 15	Hugging Face	Open New PEFT methods
llama.cpp b10994 Sep 16 ~3:27 AM ET	ggml	Nightly Metal mul_mm_id NaN
gpt-6-astra Standing	OpenAI	Cloud No new flagship id
claude-fable-5-1 Standing	Anthropic	Cloud Compaction beta only

Security / policy / access

Item	Beat	Status
AgentCore Consent portal Sep 14	OAuth session binding for Gateway	Active
RL fine-tuning console Sep 15 Preview	Gemini on Agent Platform	Watch Pre-GA
Policy Troubleshooter agents Sep 15 RN	IAM for agent identities	Watch
Claude compaction Sep 14 beta	On-demand conversation summary	Beta

Industry Shorts

Critical · Metal MoE NaN

If Metal MoE prefills (≥32 tokens) return an all-NaN vocabulary while short prompts work, pull llama.cpp ≥b10994. Author reproduction is M2 Max — verify on your silicon.

Watch · RLFT Pre-GA

Console RL fine-tuning for Gemini is Preview with Pre-GA terms — Google’s banner says no proprietary/production use. Treat as eval only until GA.

Opportunity · Ollama 0.34.1

Upgrade from 0.34.0 / drop the RC: MLX safetensors create is GA, and MLX memory handling is the Mac multi-model story to re-test on Studio.

Try This

Upgrade Ollama to v0.34.1 stable

Install GA 0.34.1 (not 0.34.2-rc0). On Apple Silicon, exercise MLX ollama create from safetensors and multi-model load memory. Confirm ollama --version reports 0.34.1.

ollama mlx 0.34.1

Metal MoE NaN → llama.cpp ≥b10994

If a high-activation MoE (release cites Mistral Small 4 / mistral4) NaNs on Metal prefill ≥32 tokens, pull macOS arm64 nightly b10994+ and re-run. Do not invent tok/s from the M2 Max note.

llama.cpp metal b10994

Issue 18 · Wednesday, September 16, 2026 · Not investment advice · Sources cited inline · Other news in AI: none this window · MLX GA quiet · Issue 17 Apigee/CVP/Genie/b10978/mlx-vlm not re-led