v1.4.8

Install: pip install freesolo-flash==1.4.8

PyPI: https://pypi.org/project/freesolo-flash/1.4.8/

What's Changed

• Support thinking-enabled tool calls without executing reasoning text by @DavidBShan in #1593
• Merge pull request #1593 from freesolo-co/fix/thinking-tool-calls by @DavidBShan in #1594
• Fix PyPI release version and flaky bootstrap cleanup CI by @DavidBShan in #1595
• Dev by @DavidBShan in #1596

Full Changelog: v1.4.7...v1.4.8

v1.4.7

Install: pip install freesolo-flash==1.4.7

PyPI: https://pypi.org/project/freesolo-flash/1.4.7/

What's Changed

• Bridge brief identity-service outages for recently verified API keys by @DavidBShan in #1586
• Stop preemption retries and use non-preemptible CPU hosting by @DavidBShan in #1589
• Add pretrained Qwen3.5 9B Base to Flash and Modal by @DavidBShan in #1587
• Fix dev package publishing after a recent release by @DavidBShan in #1591
• Stop 9B Base GRPO rollouts at assistant-turn boundaries by @DavidBShan in #1592
• Support nested scalar unions in tool schemas by @DavidBShan in #1588
• Dev by @DavidBShan in #1590

Full Changelog: v1.4.6...v1.4.7

v1.4.6

Install: pip install freesolo-flash==1.4.6

PyPI: https://pypi.org/project/freesolo-flash/1.4.6/

What's Changed

• Render SFT rows with their tool declarations by @DavidBShan in #1582
• Stop the plane when a lease takeover cannot start its background work by @DavidBShan in #1581
• Dev by @DavidBShan in #1583
• Release 1.4.6 by @DavidBShan in #1584
• Dev by @DavidBShan in #1585

Full Changelog: v1.4.5...v1.4.6

v1.4.5

Install: pip install freesolo-flash==1.4.5

PyPI: https://pypi.org/project/freesolo-flash/1.4.5/

What's Changed

• Use native 256K context across Qwen serving and preserve request errors by @DavidBShan in #1577
• Hold a control-plane lease so planes can overlap during deploys (1.4.5) by @DavidBShan in #1578
• Fix/control plane lease by @DavidBShan in #1580
• Dev by @DavidBShan in #1579

Full Changelog: v1.4.4...v1.4.5

v1.4.4

Install: pip install freesolo-flash==1.4.4

PyPI: https://pypi.org/project/freesolo-flash/1.4.4/

What's Changed

• Serve Qwen3.5-9B on H100 instead of B200 by @DavidBShan in #1571
• Reduce hosted serving GPU costs across the model catalog by @DavidBShan in #1572
• Dev by @DavidBShan in #1573
• fix(release): bump flash to 1.4.4 to unblock PyPI publishing by @DavidBShan in #1575
• Report missing SFT outputs separately from truncation by @DavidBShan in #1574
• Dev by @DavidBShan in #1576

Full Changelog: v1.4.3...v1.4.4

v1.4.3

Install: pip install freesolo-flash==1.4.3

PyPI: https://pypi.org/project/freesolo-flash/1.4.3/

What's Changed

• Preserve request credentials across Flash recovery by @DavidBShan in #1565
• Merge pull request #1565 from freesolo-co/fix/checkpoint-deploy-recovery by @DavidBShan in #1566
• fix: prevent submission lockout after early validation errors by @DavidBShan in #1567
• Rebuild image SFT pixel tensors lazily in the training dataset by @DavidBShan in #1569
• feat(custom): Tinker-style soft targets and custom training for every model by @DavidBShan in #1563
• Dev by @DavidBShan in #1570

Full Changelog: v1.4.2...v1.4.3

v1.4.2

Install: pip install freesolo-flash==1.4.2

PyPI: https://pypi.org/project/freesolo-flash/1.4.2/

What's Changed

• feat: expose top-k log probabilities in chat client and CLI by @DavidBShan in #1559
• Merge pull request #1559 from freesolo-co/feat/modal-top-logprobs by @DavidBShan in #1560
• Bump Flash to 1.4.2 to unblock PyPI publishing by @DavidBShan in #1561
• Dev by @DavidBShan in #1562

Full Changelog: v1.4.1...v1.4.2

v1.4.1

Install: pip install freesolo-flash==1.4.1

PyPI: https://pypi.org/project/freesolo-flash/1.4.1/

What's Changed

• Fix rejection of training files over 32 MiB by @DavidBShan in #1549
• Merge pull request #1549 from freesolo-co/fix/large-training-files by @DavidBShan in #1550
• Expose GRPO rollout token probabilities to reward functions by @DavidBShan in #1551
• Add authenticated checkpoint adapter downloads by @DavidBShan in #1552
• Fix package release version and rate-limited test notifications by @DavidBShan in #1555
• Size the image SFT profile decode budget for datasets by @DavidBShan in #1557
• Replace GitHub managed environment storage with Azure Blob by @DavidBShan in #1554
• Remove built-in OPD and managed teacher infrastructure by @DavidBShan in #1556
• feat: add managed custom algorithms with durable resume and async control by @DavidBShan in #1553
• Dev by @DavidBShan in #1558

Full Changelog: v1.3.33...v1.4.1

v1.3.33

Install: pip install freesolo-flash==1.3.33

PyPI: https://pypi.org/project/freesolo-flash/1.3.33/

What's Changed

• fix: request JSON in JSON-constrained deployment smoke by @DavidBShan in #1544
• Merge pull request #1544 from freesolo-co/fix/grpo-serving-smoke by @DavidBShan in #1546
• Fix multi-turn screenshots, auth outages, and serving release promotion by @DavidBShan in #1543
• Make GRPO truncation masking explicit and remove the stop-string workaround by @DavidBShan in #1547
• Dev by @DavidBShan in #1548

Full Changelog: v1.3.32...v1.3.33

v1.3.32

Install: pip install freesolo-flash==1.3.32

PyPI: https://pypi.org/project/freesolo-flash/1.3.32/

What's Changed

• Fix warm-start GRPO crash on meta-initialized FSDP ranks by @DavidBShan in #1539
• Fix warm-start GRPO initialization on meta FSDP ranks by @DavidBShan in #1540
• Dev by @DavidBShan in #1542

Full Changelog: v1.3.31...v1.3.32