SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Developer fine-tunes 27B model to write Atlassian Forge apps, releases weights openly

0
·1 views

A developer has fine-tuned the Qwen3.8-27B language model to generate functional Atlassian Forge applications, releasing the weights publicly on Hugging Face under the Apache-2.0 license. The base model previously passed only 15% of Atlassian identifier probes and produced zero valid Forge manifests out of 25 attempts, but after fine-tuning those figures rose to 69% and 14 of 25 respectively. Training was conducted as a rank-32 LoRA over approximately 26 hours on a single Apple Mac Studio M3 Ultra with 96 GB of memory. The model was trained on real Forge apps, official documentation, OpenAPI specs, and Atlassian Community answers covering Jira, Confluence, and Jira Service Management. The developer claims this is the first open-weights model specifically trained to write Forge apps and the first Atlassian-tuned release accompanied by base-versus-tuned validation evidence.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

LLM Routing Tools Compared in 2026: Bifrost Leads With 11-Microsecond Overhead

Model routing tools have evolved to dynamically direct AI inference requests across multiple LLM providers based on cost, latency, and provider health. Production applications that send all prompts to a single frontier model typically overspend by 40–80% on simple queries while remaining vulnerable to provider outages and rate limits. Bifrost, an open-source AI gateway built in Go by Maxim AI, tops the 2026 rankings by delivering sub-millisecond gateway overhead of just 11 microseconds at 5,000 requests per second. Other notable tools include open-source proxy LiteLLM, classifier-based RouteLLM, and hosted aggregator OpenRouter, each addressing different routing needs with distinct trade-offs. Engineering teams evaluating routing platforms are advised to assess six key dimensions: routing logic expressiveness, runtime latency, failover capability, provider coverage, governance controls, and deployment flexibility.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Ex-OpenAI Researcher Launches Jev, a Decision-Only AI Model That Cannot Generate Text

Diogo Almeida, the OpenAI researcher behind RLHF and InstructGPT, has launched Jev through his startup TypeSafe AI, releasing it in early access on 15 September. Unlike large language models, Jev is designed solely to make structured decisions, returning one of three output types — a Choice, a Score, or a probability — each paired with a calibrated confidence value. The model completes tasks in 70 to 500 milliseconds and charges $0.042 per million input tokens, with no fee for output due to its non-autoregressive architecture. Almeida argues that chat was the wrong target for advancing AI automation, and that reliable, fast decision-making with trustworthy confidence scores is what software pipelines actually need. TypeSafe AI raised $40 million in funding and built the team from alumni of OpenAI, Google Brain, Meta, and FAIR, with Jev's benchmarks claiming it runs up to 193x faster and 444x cheaper than frontier language models on comparable workflows.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Poor database indexing caused 'not found' API queries to take up to 30 seconds

A maritime tracking API was found to be dramatically slower when returning empty results than when returning actual data, a problem traced to misconfigured TimescaleDB compression settings. On September 9, 2026, queries for a vessel's latest port event took 28 milliseconds when data existed but 2 to 2.5 seconds when it did not, while ETA lookups for unmapped vessels could stall for 30 seconds or more. Over a single night, 'not found' requests consumed 71% of all database time spent on the ETA endpoint. The root cause was that compressed data chunks lacked usable indexes on the columns being queried, forcing the database to perform a sequential scan across hundreds of compressed chunks for every miss. Grouping API metrics by route template rather than raw path revealed the pattern, showing that slow responses were consistently tied to empty results rather than specific vessels or endpoints.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Critical MikroTik RouterOS Flaws Allowed Full Takeover Without Credentials

CERT Polska disclosed a two-vulnerability exploit chain called MikroTrick in September 2026, targeting MikroTik RouterOS devices reachable via SSH over the internet. The first flaw, CVE-2026-67276, is an SSH public-key authentication bypass caused by a logic error that ignores the RSA exponent during key verification, allowing unauthenticated access. The second, CVE-2026-86060, is a privilege escalation triggered by a malformed username, granting full administrative control once an attacker is authenticated. Both vulnerabilities carry critical CVSS scores between 9.2 and 9.8, and active exploitation was recorded from September 2, 2026 — one day before MikroTik released patched versions across its 6.x and 7.x release trains. Unpatched devices exposed to the internet should be treated as potentially compromised, and administrators are advised to upgrade immediately or restrict SSH access to trusted addresses as an interim measure.

0 comments Read more at DEV Community