v0.3.2 — 9B 17/17 Measured, Security Fixes, igllama v0.3.11

Highlights

9B Achieves 17/17 at All Quantizations

The full 9B T01-T17 trial is complete. Every quantization variant (Q4, Q5, Q6, Q8, BF16) passes all 17 agentic tasks — code generation, JSON round-trip, error recovery, multi-source synthesis, everything. 9B is the gold standard for local agentic tool use.

9B Variant	Score	Turns	Time
9B-Q4	17/17	38	9127s
9B-Q5	17/17	49	14239s
9B-Q6	17/17	39	8642s
9B-Q8	17/17	39	17324s
9B-BF16	17/17	43	15722s

Security and Correctness Fixes

• OAIResponse use-after-free — send() now dupes all string fields before returning (previously returned dangling pointers into freed JSON parse tree)
• Config save malformed JSON — save() rewrote using std.fmt instead of broken manual string concatenation
• getToolCalls page_allocator leak — replaced leaking ArrayList with proper allocator-based allocation
• Auth header memory leak — Bearer {key} string now freed after HTTP request

igllama v0.3.11

• Strip residual </think> tokens when --no-think is active (both streaming and non-streaming paths)

Verified

• 195/195 tests pass, 0 leaks
• GH Pages rebuilt and deployed

v0.3.1 — Context Sensitivity Harness + Trial Filter + 4B T01-T17 Measured

What's New

Context Length Sensitivity Harness

New examples/ctx_sensitivity.zig measures 2B-Q6 accuracy across ctx-size 512/1024/2048/4096 × T01-T17. Run with zig build ctx.

Trial Quantization Filter

trial-quant now accepts optional model names as CLI arguments:

./zig-out/bin/trial-quant 9B-Q4 9B-Q5     # only these two
./zig-out/bin/trial-quant                  # all 16 models

4B-Q4 T01-T17 Fully Measured

• Result: 15/17, 63 turns, ~9050s
• T04 (multi-step write) and T16 (Zig compile recovery) fail via turn exhaustion at 1.3 tok/s
• These are characteristic failure modes for 4B at Q4 — not correctness failures

Verified

• 195/195 tests pass, 0 leaks
• 9B T01-T17 re-run in progress (background); showcase will be updated when complete

v0.3.0 — 4B Quant Curve, Throughput Benchmark, Extended Task Suite

What's New

New: Throughput Benchmark (examples/bench.zig)

zig build bench

Measures tokens/second for each Qwen3.5 model across Q4/Q8/BF16 precision levels. Reports tok/s, file size, and RAM (RSS). Requires igllama v0.3.10+ for accurate usage.completion_tokens.

Measured results (CPU-only, 4 threads):

Model	tok/s	RAM
0.8B-Q8	3.4	0.8 GB
0.8B-BF16	2.9	1.5 GB
2B-Q4	2.9	1.3 GB
2B-Q8	2.6	1.9 GB
2B-BF16	1.9	3.6 GB
4B-Q4	1.3	2.7 GB
4B-Q8	0.1	~4 GB (swap!)

Key finding: RAM is the hard limit. Models that exceed physical RAM fall off a cliff (4B-Q8: 0.1 tok/s from swap thrashing). 4B-Q4 is the practical ceiling on systems with ≤6 GB free RAM.

Extended: 4B Full Quant Curve

Added Qwen3.5-4B-Q4_K_M, Q5_K_M, and Q6_K to trial_quant.zig. All three pass 13/17 — 4B is saturated at Q4. 4B-Q4 (2.6 GB) is the recommended production config: full accuracy, minimum file size.

Extended: T01–T17 in trial_quant.zig

The quantization sensitivity harness now runs all 17 tasks (was T01–T13), adding code generation with zig fmt validation (T14), JSON round-trip (T15), error recovery (T16), and multi-source synthesis (T17) across all 16 quantization variants.

trial_quant.zig now covers: 16 models × 17 tasks = 272 test cases per full run.

igllama v0.3.10 Upstream Fix

Patched usage.completion_tokens in igllama's non-streaming /v1/chat/completions handler — was hardcoded 0, now returns real counts. Fix upstreamed as igllama PR #82, released as igllama v0.3.10.

Upgrade Notes

• Requires igllama v0.3.10+ for accurate bench token counts (fallback estimate still works with older builds)
• Download 4B quant GGUFs: igllama pull unsloth/Qwen3.5-4B-GGUF -f Qwen3.5-4B-Q4_K_M.gguf

Stats

• 195/195 tests pass
• 3 harness executables: trial, trial-quant, bench
• 16 models × 17 tasks in quantization harness

v0.2.7 — BF16 precision trials, igllama grammar fix, LICENSE

What's new in v0.2.7

Added

• BF16 in trial_quant.zig — 2B-BF16 and 9B-BF16 added to QUANT_MODELS; harness now covers the full Q4/Q5/Q6/Q8/BF16 precision curve
• LICENSE — MIT license file added

Fixed

• igllama v0.3.10 — streaming json_mode use-after-free — streaming handler freed the grammar string while the sampler held a pointer to it; replaced with direct JSON_GRAMMAR comptime const (matches non-streaming handler)
• trial_quant.zig — changed response_format from json_object to text; the grammar sampler in the bundled llama.cpp crashes during generation for 2B+ model vocabularies; system prompt JSON constraint is sufficient for 4B+

Changed

• showcase.smd — documents igllama json_mode crash finding, expanded trial task suite to T01–T17, updated framework version

Build

zig build trial-quant   # Q4/Q5/Q6/Q8/BF16 sensitivity trial for 2B + 9B
zig build trial         # T01-T17 across all 4 weight classes
zig build test          # 195/195 tests

v0.2.2 — Session summary, igllama port scan, json_mode, Showcase

What's new in v0.2.2

Features

• Session summary output — powerglide run now emits a structured completion block with steps, elapsed time, agent/model, and the <POWERGLIDE_DONE> or <POWERGLIDE_ERROR> terminal signal
• igllama port scan — powerglide doctor scans :8090–8099 and reports all running igllama instances simultaneously
• json_mode on OpenAIClient — sets response_format: {"type":"json_object"} to force constrained JSON output from igllama and other local endpoints

New Showcase page

Live at bkataru.github.io/powerglide/showcase — four case studies documenting powerglide dogfooding with Qwen3.5 0.8B and 4B models via igllama, including the honest tool calling triage and performance table.

Bug fix

• Loop step count increments test now uses an isolated /tmp session file; previously picked up real .powerglide/session.json from dogfooding runs

Stats

• 195/195 tests passing, 0 memory leaks
• Fully local stack: powerglide + igllama + Qwen3.5-4B, no API keys required

v0.2.1 — 195 tests, bug fixes

v0.2.1

Test Coverage Expansion

• 195/195 tests passing (up from 170)
• New test modules: SSE parser, HTTP response, persistence manager
• Root module now covers all submodules via refAllDecls

Bug Fixes (uncovered by expanded coverage)

• stream.zig: unmanaged ArrayList API fixes (Zig 0.15.2 compliance)
• terminal/pool.zig: sessions.size → sessions.count()
• terminal/session.zig: array literal syntax fix, orphaned test code removed

195/195 tests, 0 leaked.

v0.2.0 — MCP Integration

What's New in v0.2.0

MCP Integration

• MCP Server — powerglide mcp starts powerglide as a JSON-RPC 2.0 MCP server over stdin/stdout, exposing all registered tools to any MCP-compatible client
• MCP Client — connect to external MCP servers; their tools become first-class powerglide tools prefixed as mcp_{server}_{tool}
• Tool Bridge — transparent McpTool → Tool conversion for seamless integration
• Config support — mcp_servers array in ~/.config/powerglide/config.json

Fixes

• Stdin API fix for Zig 0.15.2 (posix.read byte-by-byte pattern)
• Favicon 404 on GitHub Pages resolved
• Homepage title deduplication

138/138 tests passing.