Observed Signal · Aug 28, 2026 · Technical Release · Source: DEV Community · Impact: 2/5 · Sentiment: Positive
Flash Onyx 2.2: Local Model for Law and Game Feel
Flash Onyx 2.2 is a development update for a local-first agent shell (Flash Onyx) built around gemma4 configured as an engineering agent that runs fully on a 16 GB laptop GPU at single-digit tokens per second. Version 2.2 expands the system-prompt domains beyond code into legal and game-design guidance, with strict rules (e.g., refuse invented citations, jurisdiction-first for law; game feel heuristics for games). The author compressed prompt text to reduce per-conversation prefill cost, tuned sampling (temperature moved from 0.7 to 0.6 and min_p from 0 to 0.05), and increased context window to 65536 after probes showed 32k/64k/128k used ~8.1–8.2 GB GPU memory. Build tooling now injects the repo license into models; Onyx 2.2 is not yet published but is planned as Natuworkguy/flash-onyx-2.2 in 12b and 31b variants.
Practical technical guidance on prompt design, sampling, and context-window tuning for local LLMs is useful to developers building private/local agents but is a niche developer-level update, not industry-shifting.
Track GitHub Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your Polaris7 workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Flash Onyx 2 is based on gemma4 configured as an engineering agent.
- The model runs locally on a 16 GB laptop GPU at single-digit tokens per second.
- Version 2.2 adds focused system-prompt domains for LAW and GAMES with explicit rules (e.g., do not invent legal citations; prioritize feel in games).
- Sampling was tuned: temperature lowered from 0.7 to 0.6 and min_p changed from 0 to 0.05.
- Context window was set to 65536 after probes showed 32k/64k/128k windows used ~8.1–8.2 GB GPU memory; prefill rate is ~95 tokens/second.
Connected Companies & Entities
“Flash Onyx is the model line behind FLASH (https://github.com/natuworkguy), the local-first agent shell I work on.”
“The build tooling grew up a little too. models/build.py fills the model name into the prompt from the header, and injects the repo license i...”
“At 0.5 it also stopped, but reached for du --max-depth=1, a GNU flag this Mac does not have, breaking a different rule in the same prompt.”
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Flash Onyx 2.2 Released, Finishes Model Improvements
Flash Onyx 2.2, a new version of the Flash Onyx model by the author 'Natuworkguy', is published and available via Ollama. The release ships in two sizes (12b for consumer hardware and 31b for GPU-equipped machines) and focuses on improving request comprehension, producing correct concise answers, automatic stack selection, safer and more accurate Manim animation code generation, and better game-development patterns. Sampling and context settings were adjusted (temperature 0.7→0.6, min_p 0.0→0.05, num_ctx 32768→65536). The post includes install/pull commands for Ollama, recommended MODEL env settings, and an install script link. The author invites feedback and notes that future work (2.3) will address any failures encountered when using 2.2.
Google Unveils Gemini 3.5 Flash and Antigravity 2.0
At Google I/O 2026 Google announced Gemini 3.5 Flash — a model optimized for high-speed, low-latency agent execution — and Antigravity 2.0, a standalone desktop application for coordinating multiple autonomous agents in parallel. Gemini 3.5 Flash is presented as an engine tuned for recursive agentic workflows, quoted at 289 output tokens per second and reported to outperform earlier flagship models on agent-focused benchmarks (Terminal‑Bench 2.1 and MCP Atlas). Antigravity 2.0 acts as a command center that decouples agent conversations from a single repository, enables true parallel multitasking (specialist agents working concurrently), and can manage separate local Git branches without overlap. The announcements frame a shift from inline autocomplete toward agent orchestration and raise practical concerns about verification, auditing, and developer roles.
Gemma 4 Enables Local Multimodal, Long-Context Workflows
A developer reports replacing fragmented OCR + RAG stacks with local Gemma 4 models, claiming the model family makes coherent, private, on-device multimodal intelligence practical on consumer hardware. Using the Ollama Python SDK and local inference, the author says Gemma 4’s 26B MoE and 31B Dense variants reason over pixel layouts directly (no separate OCR), achieving ~94% extraction accuracy on complex receipts with simple image preprocessing on an M1 MacBook Pro (16GB). Gemma 4’s native 128K context window allowed the author to ingest a continuous 115K-token log stream and trace a multi-month causal chain in ~70 seconds, highlighting temporal coherence benefits over chunked RAG. The post lists recommended model/context budgets, notes limits (very degraded inputs, real-time latency, knowledge cutoffs), and cites Gemma developer docs and Ollama resources. Publication date: 2026-05-21.
Receive all future signals for GitHub & competitors in your briefing
Never miss material M&A deals, leadership changes, account wins, or product rollouts. Polaris7 monitors 20,000+ primary sources 24/7, delivering verified telemetry directly into your executive briefing.